WO2015189562A1 - Image capture apparatus and method - Google Patents

Abstract

An imaging apparatus and method are provided configured for discriminating between sky and non-sky regions. The apparatus comprises a UV-passing image forming element and a UV-sensing image capturing sensor for capturing an image projected from the UV- passing image forming element. The image comprising at least one of sky and non-sky image regions and the image captured by the sensor is provided to an image processing unit for performing discrimination between sky and non-sky regions of the captured image using primarily UV wavelengths.

Description

IMAGE CAPTURE APPARATUS AND METHOD

FIELD OF THE INVENTION

The present invention relates to an image capturing apparatus and method. In particular, the present invention relates to an imaging system for discriminating between sky and non- sky portions of an image, for example, in a sky panorama.

BACKGROUND OF THE INVENTION

Navigation systems such as those relying upon Global Positioning Satellites, High Definition (HD) cameras, RADAR, laser based Light Detection and Ranging (LIDAR) scanners and Inertia Measurement Units (IMU) as sensor technologies are known and have been deployed to enable robotic vehicles to perform at least partially autonomous navigation. However, these known sensor technologies can be expensive to manufacture and power-hungry. Thus there is a need for more cost-effective and low-power navigation systems for wider deployment in robotic devices.

It is known that a silhouette of sky against terrestrial objects such as buildings and mountains can be used to aid localisation by allowing recognition of a unique location. Indeed "sky segmentation", which involves categorising image pixels as sky or non-sky, can be used to achieve more accurate position information than GPS in locations such as cities where GPS is typically degraded by the presence of high-rise buildings. Sky segmentation using visible light is known, but obtaining reliable results under changing weather conditions and ambient light levels can be problematic, meaning that complex machine-learning algorithms have to be employed to make the visible-light systems more robust. Other known imaging systems use deep Infra-red (I R) light instead of visible light or use a combination of wavelengths using two different "channels" such as Ultra-Violet (UV) and IR or UV and green wavelengths to obtain results that are less variable to differing overall light intensity. The Infra-red range of the electromagnetic spectrum is nominally 700nm up to 1 mm wavelengths. The UV range of the electromagnetic spectrum is approximately 100nm up to 400nm, although the Earth's atmosphere is opaque to wavelengths less than about 200nm. The use of two different "channels" in this way has been inspired by the way that desert ant eyes are proposed to function. The dual channel methods seek to use the dual channels to find a threshold for each individual pixel, because desert ants do not integrate light across the full image. These known techniques focus upon finding an optimal combination of two bandwidths to achieve robustness to changing light intensity in the image-capture environment.

Thus there is a requirement for an image capturing system for performing sky segmentation that is more resilient to changes in light intensity and changing weather conditions in the image-capturing environments and which is cheaper to manufacture and simpler to implement than previously-known systems. Such image capturing systems have applications in fields such as robotics, but are also applicable to other fields such as, for example, plant canopy measurement.

BRIEF SUM MARY OF THE DISCLOSURE

According to a first aspect, the present invention provides an imaging apparatus configured to discriminate between sky and non-sky, the apparatus comprising:

a UV-passing image forming element;

a UV-sensing image capturing element for capturing an image projected from the UV-passing image forming element, the image comprising at least one of sky and non-sky image regions;

an image processing unit for performing discrimination between sky and non-sky - regions of the captured image using primarily UV wavelengths.

The image regions could be, for example, image elements, pixels or groups of pixels. The use of primarily UV wavelengths, for example, a subset of UV wavelengths corresponding to 320nm to 400nm, has been demonstrated to be robust to changes in light levels, illumination angles and weather conditions in a way that the use of visible and/or I R wavelengths is not. Furthermore, the system is simpler to implement and more efficient than previously known systems that use combinations of wavelength bands such as UV and green to perform sky segmentation. Due to a single wavelength band (UV) being used, a single-value intensity metric can be used to generate a binary sky versus non-sky image pixel profile. Thus the requirements for post-processing of the image data to generate the sky-segmented image are simplified. The primarily UV wavelengths include mainly but not necessarily exclusively UV wavelengths. For example a minority visible light or IR wavelength component could be included in the capture image upon which the discrimination between sky and non-sky regions is performed. The primarily UV wavelengths comprise a at least one subset or even a plurality of subsets of UV wavelength bandwidths within the full UV wavelength range.

In some embodiments the image processing unit is configured to perform an intensity categorisation of the image elements of the captured image, the intensity categorisation being adaptive to image characteristics comprising at least one of: light levels; image capture time; weather conditions; and illumination angle.

In some embodiments the intensity categorisation performed by the image processing unit uses a single-dimension intensity metric comprising primarily UV wavelengths. In some embodiments the image processing unit is configured to perform the discrimination between sky and non-sky image elements using a subset of primarily UV wavelengths selected from the UV wavelength range of 200nm to 400nm.

In some embodiments the UV-passing image forming element comprises at least one of a substantially hemispherical lens and a fish-eye. In other embodiments the UV- passing image forming element comprises at least one curved mirror.

In some embodiments the UV-passing image forming element comprises an array of image forming elements arranged on a curved surface, each image forming element having a different field of view such that together the array of image forming elements is configured to enable substantially simultaneous capture of a sky panorama.

In some embodiments each of the image-forming elements of the array of image- forming elements has at least one of: a respective UV-sensing image capturing element and a respective UV-passing filter forming an integral unit with the respective image- forming element mounted on the curved surface.

In some embodiments the UV-passing image forming element and the UV- capturing image forming element are disposed together as an image-capturing unit on a rotatable mounting such that a sky panorama is obtained by rotating the image-capturing unit to obtain a composite image of a solar hemisphere.

In some embodiments the UV-sensing image capturing element is configured to be sensitive primarily to UV light.

In some embodiments the image processing unit is configured to receive greyscale image data from the image-capturing sensor and to bin the greyscale image data according to pixel intensity values to obtain binary image data, the binary values discriminating between pixels associated with sky and pixels associated with non-sky.

In some embodiments the image processing unit is configured to establish a threshold intensity value from the greyscale image data such that pixels having an intensity value below the threshold are identified as non-sky pixels and pixels having an intensity value above the threshold are identified as sky pixels.

In some embodiments the threshold intensity value is determined by identifying in a histogram of the binary image data an intensity peak associated with sky pixels and an intensity peak associated with non-sky pixels and selecting the threshold intensity value to be an intensity located between the sky peak and the non-sky peak.

In some embodiments a centre of gravity of pixels corresponding to sky in the binary image data is used to perform tilt correction in a navigation system.

According to a second aspect a robot navigation system comprising the image capturing apparatus as described above is provided.

According to a third aspect, the present invention provides a method of image capture comprising the steps of:

directing light from a sky panorama through a UV-passing image-forming element; capturing an image projected from the UV-passing image forming element using an image-capture sensor, the image comprising at least one of sky and non-sky image regions;

performing discrimination between sky and non-sky regions of the captured image by performing processing of pixels of the captured image using primarily UV wavelength components of the captured image.

BRIEF DESCRIPTION OF THE DRAWINGS

Embodiments of the invention are further described hereinafter with reference to the accompanying drawings, in which:

Figure 1 schematically illustrates an imaging apparatus including a curved lens for capturing an image for sky segmentation;

Figure 2 schematically illustrates an imaging apparatus including a curved mirror instead of a curved lens;

Figure 3 schematically illustrates an imaging apparatus including a pair of curved mirrors instead of a curved lens;

Figure 4 schematically illustrates an imaging apparatus comprising an array of self-contained imaging units each comprising a lens and a respective image-capture sensor, all mounted on a curved surface;

Figure 5 schematically illustrates a rotatable image capturing unit comprising a lens, a UV-passing filter and an image capturing sensor;

Figure 6 is a flow chart schematically illustrating an algorithm used to perform sky segmentation based upon greyscale image data from the imaging apparatus of any one of Figures 1 to 5; Figure 7 is a graph of pixel frequency against intensity for a greyscale image captured by the imaging device of any one of Figures 1 to 5;

Figure 8a is an image capture apparatus comprising a camera and parabolic mirror used to test the present technique;

Figure 8b schematically illustrates pre-processed images used by navigation algorithms showing how tilt is removed;

Figure 9 is a table illustrating a relative accuracy of sky segmentation for a range of different locations;

Figure 10 schematically illustrates a comparison of sky segmentation in visible light, using UV-G opponency and according to the present technique using primarily UV wavelengths;

Figures 1 1 a to 1 1 c schematically illustrates the reliability of sky segmentation using primarily UV wavelengths across different weather conditions; and

Figure 12 schematically illustrates results of using sky segmentation to navigate in a city landscape.

DETAILED DESCRIPTION

Figure 1 schematically illustrates an imaging apparatus according to an embodiment of the invention. The arrangement comprises an image capturing module 100 including a hemispherical lens 1 10, an ultra-violet (UV) passing filter 120 and a charged-coupled- device (CCD) array 130. Image data from the image capturing module 100 is then supplied to a general purpose computing device 150, which executes program code for performing image processing. In this embodiment black and white digital image data output by the CCD array 130 is processed using a sky segmentation algorithm 160, described in detail below. A general purpose computing device is shown in the Figure 1 embodiment, for performing processing of the captured image data using general purpose processor circuitry configured by program code to perform specified processing functions. However, the image processing could alternatively be performed by a Digital Signal Processor, by dedicated special purpose hardware (e.g. a System on Chip or Application Specific Integrated Circuit). The image processing functions may be performed entirely in hardware, entirely in software of using a combination of hardware and software.

The hemispherical lens 1 10 of Figure 1 is a wide angle lens allowing the entire solar hemisphere to be viewed so that the entire sky can be sampled and captured. In alternative embodiments, the hemispherical lens is replaced by a "fish-eye" type lens having multiple lenses configured to allow substantially simultaneous viewing of an entire hemisphere. Light visible to the human eye spans the wavelength range of approximately 400 nanometres (nm) to 750 nm. Ultraviolet light has wavelengths shorter than the wavelengths of visible light and thus span the range from 1 nm up to about 400nm. However, air is opaque to wavelengths of less than about 200nm, so the UV part of the spectrum that can potentially be captured by the imaging apparatus of Figure 1 is in the range of 200nm to 400nm. However, the lens material of the hemispherical length has an effect on which UV wavelengths can pass through to the CCD array 130. Not all lenses will pass UV light, in fact most types of glass will allow longer-wavelength UV light to pass, but will absorb UV wavelengths of about 350nm and below. Accordingly, the hemispherical lens 1 10 is a UV-passing lens specially configured to allow more UV wavelengths to pass through than standard lenses. UV-passing lenses such as those made from quartz or quartz and fluorite may be used for this purpose. The image-capturing element according to some embodiments may capture wavelengths outside the UV spectrum, provided that it a substantial UV component is captured. The primarily-UV component of the image may be isolated via filtering in the image capture arrangement itself or could alternatively be isolated, for example, by post-processing of the captured image using post processing hardware and/or software.

The UV- passing filter 120 in the embodiment of Figure 1 allows ultraviolet light to pass but blocks transmission of most visible and infrared light. The UV-passing filter 120 may comprise a UV-transmission filter may itself block infrared light, or alternatively, may be combined with together an additional infrared blocking filter.

Light from all angles of the sky passes through the hemispherical lens 1 10 and this lens focuses light onto the CCD array 130 via the UV-passing filter 120. The CCD array 130 transforms light incident on each CCD pixel into electrons. Typically the electrical charge is transported across a CCD chip and read at one side of the array. An analogue-to-digital conversion is performed so that a digital intensity value is obtained from the CCD array 130 for each pixel. Thus the CCD array 130 forms a greyscale pixelated image of the sky based upon detection of primarily a UV wavelength-range of light from a full panorama of the sky (or composite image of the solar hemisphere). The greyscale digital image data from the CCD array 130 is then analysed by one or more algorithms on the computer 150 to perform a sky segmentation process, whereby each pixel of an image is assigned a binary value of either "sky" or "non-sky" based on the detected pixel intensity values. In alternative embodiments the CCD array 130 is replaced by a complementary metal-oxide semiconductor (CMOS) array. The CCD array 130 will typically produce a higher-quality, lower-noise image that a CMOS array, but will consume more power and be more expensive to manufacture.

In yet further alternative embodiments, the CCD array 130 could be configured to detect light in a broader range of wavelengths that primarily UV light (e.g. including at least some of the visible spectrum) and the portion of the spectrum other that the UV portion could be filtered out of the image during a post-processing stage before the pixels are categorised as sky or non-sky. In one embodiment the primarily UV portion of the electromagnetic spectrum used to perform the sky segmentation is an incomplete portion of the full UV wavelength range, such as the range of 320nm to 400nm, which is closest to the visible spectrum. However in other embodiments a full 200nm to 400nm UV wavelength range or a 200nm to 300nm wavelength range could be used as alternatives.

Figure 2 schematically illustrates an alternative image capturing apparatus according to the present technique, in which a curved mirror is used to capture the sky panorama. The curved mirror is a parabolic mirror in this embodiment but in alternative embodiments could be, for example, a hyperbolic or conical mirror. The arrangements of Figure 2 and Figure 3 can be used to sample a full hemisphere, not merely a semi-circular slice of the sky. The arrangement comprises a curved mirror 210, a UV-passing filter 220, a regular lens 230 and a CCD/CMOS detector array 240. In this particular embodiment, the CCD/CMOS detector array 240, the regular lens 230, and the UV filter are all mounted on a Digital Single Lens Reflex (DSLR) camera 260 as part of a camera-mounted assembly. The curved mirror 210 is located sufficiently far from the camera assembly to collect light from the sky panorama and direct the collected light onto the CCD/CMOS detector array 240. Light reaches the convex face of the curved mirror 210 through the side wall of a UV- passing tube 250, which fixes the curved mirror 210 in a substantially fixed alignment with the CCD/CMOS detector array 240, allowing the skyline panorama image to be captured. The UV-passing tube 250, similarly to the UV-passing lens 1 10 of the Figure 1 embodiment, is made of a quartz or quartz-composite material. However, any UV-passing material could be used.

Figure 3 schematically illustrates an alternative embodiment of an imaging apparatus in which a pair of curved mirrors is used to direct a sky panorama image onto a CCD array 340. This arrangement comprises a convex mirror 305 and a concave mirror, both arranged with their concave surface facing the CCD array. Similarly to the arrangement of Figure 2, a UV-passing filter 320 is mounted on top of a regular lens 330 on top of the CCD array 340 on the end of a DSLR camera 360. A UV-passing tube 370 connects the two mirrors 305, 310 to the camera assembly 320, 330, 340, 360. The convex mirror 305 sits closer to the CCD array 340 than the concave mirror 310. Light of a broad range of wavelength including visible light passes through the cylindrical wall of the UV-passing tube 370, and is reflected from the convex mirror 305 onto the concave mirror 310, where incident light is reflected back towards the mirror 305 and through an aperture at the apex of that mirror 305 so that it is directed towards the CCD array 340.

Figure 4 shows a cross-section through an array of image capturing elements according to a third embodiment. In this embodiment an array of image forming elements is arranged on a curved surface. This array of image-forming elements is used to sample a full sky hemisphere. The arrangement comprises an array of micro-lenses 410 mounted on the surface of a hemisphere. The lenses in this embodiment are regular lenses rather than hemispherical lenses. A complementary array of photoreceptors 420 (for example a CCD or CMOS sensor) is arranged on an inner concentric inner hemispherical surface such that light from each individual micro-lens 410 is focused on a respective photoreceptor 420. -

In Figure 4 a single UV-passing filter 430 that spans the entire lens array is provided, but in alternative arrangements a separate UV-passing filter is provided for each microlens such that each image forming element of the array is an encapsulated unit comprising a UV- passing filter, a lens and a photoreceptor. I n the lens array arrangement of Figure 4, each pixel samples from a different portion of the sky panorama directly so that each encapsulated image-forming element is directed to a single portion of sky and to a small area of the entire field of view. This lens array arrangement is analogous to the compound eye of an insect. In yet further alternative embodiments of the lens array arrangement, a plurality of lenses are served by a single UV filter and image sensor by using an optical fibre to direct light from each lens to respective portions of the image sensor via a common UV-passing filter. For all of these lens array arrangements, data captured by all of the image sensors is collated by a single integrated circuit (not shown). Figure 5 is yet a further alternative arrangement in which a single encapsulated unit 510 comprising a UV-passing filter, a UV-passing lens and a light sensor element is rotated on a trajectory to obtain a composite image of the solar hemisphere. This arrangement is used to sample a full hemisphere. A plurality of these rotatable encapsulated units may be provided in alternative embodiments.

Figure 6 is a flow chart that schematically illustrates processing performed on UV image data captured from any one of the image capturing apparatuses of Figures 1 to 5. According to embodiments, an adapted "watershed" inspired algorithm, which is simple to implement and efficient to run, suffices to produce sky segmentation to analyse composite data from an entire solar hemisphere, unlike previously known techniques which employ complex machine-learning approaches or require combination of two or more distinct light bandwidths to segment the sky. The adapted watershed algorithm is only one example of how greyscale image data, generated using primarily UV wavelengths from the solar hemisphere, can be processed to distinguish between sky and non-sky regions. In some embodiments each pixel is denoted sky or non-sky. Alternatively groups of pixels or other image elements could be denoted sky or non-sky. The captured images may in some cases be substantially 100% sky or substantially 100% non-sky, although more typically a mixture of sky and non-sky regions are present.

The adapted watershed algorithm begins at process element 610 where greyscale UV detection data is received from the image sensor elements of, for example, the CCD array 130 of Figure 1. Next, at process element 620, the pixel greyscale intensity values are binned to form a histogram of number of pixels against intensity. At stage 622 a smoothing operation is performed to remove small local peaks whilst preserving more prominent features of the histogram. An example of the unbinned pixel intensity values is shown in Figure 7, where a first peak 710 at an intensity of around 1000 corresponds to a non-sky contribution, a second peak 720 at a higher intensity value of around 5000 corresponds to pixels associated with the sky, whilst a third peak 730 at yet higher intensity of around 7500 is an artefact of the sun. Returning to Figure 6, at stages 630 and 640 a horizontal line (740 in Figure 7) parallel to the intensity axis (X-axis) of the Figure 7 graph starts on the X axis and is progressively moved from the intensity-axis higher up the pixel number scale of the Y-axis. As the threshold line is raised, the number of crossing points of the line with the intensity profile should change. It can be seen from the illustrative example data of Figure 7 that initially the line will have six crossing points due to the third peak 730 (artefact of sun), but there will reach a point when the line lies above the third peak 730 and reducing the number of crossing points from six to four. This corresponds to process element 650 in Figure 6.

The position of the horizontal test-line is effectively fixed corresponding to the lowest pixel number value for which there are only four crossing points 762, 764, 766, 768. However, the adapted watershed algorithm does not search specifically for four crossing points because if the image pixels are fully saturated, the left-most peak could be peaked over zero intensity and the right-most peak could correspond to an upper bound, meaning that in the worst case only two intersection points would be detected. Instead, the adapted watershed algorithm searches for a single intersection of the horizontal test line with a positive slope (gradient) of the intensity histogram that occurs after the first (i.e. lowest intensity) intersection with a negative slope.

Thus the lowest intensity crossing point 762 on Figure 7 is disregarded because, although it corresponds to an intersection of the line by the captured greyscale image intensity profile with a positive slope, there is no negative slope intersection to the left of the intersection point 762. However, a subsequent crossing point 766 is found to correspond to a positive gradient intersection and this counts as a positive detection by the algorithm. The identification of crossing point 766 triggers the beginning of the search for a sky/ground threshold intensity (a line parallel to the Y0 axis on the histogram) because it corresponds to a single intersection of the horizontal test line (by a positive slope of the image intensity histogram) after a first intersection 764 with a negative slope. Continuing the search on the Figure 7 intensity profile, after point 766, no further intersections with a positive slope are found so it is deduced that the intersection 766 is likely to correspond to the left side of an intensity peak associated with the sky. The final intersection point 768 is ignored because it corresponds to an intersection having a negative gradient. It will be appreciated that if the horizontal test line is set to lie closer to the X-axis it may have further intersections corresponding to the right-most peak in Figure 7. In this case there would be two intersections of the test line having a positive slope that come after intersections having a negative slope. According to the present technique, the horizontal test line is raised until there is only one such combination of positive slope intersection immediately preceded by negative slope intersection. Thus process element 650 in Figure 6 involves correctly fixing the horizontal test line such that it lies at a lowest intensity value for which there is a single combination of positive gradient intersection preceded by a negative gradient intersection. At process element 660, an intensity-discriminating value is worked out as the intensity value 770 falling half way between the first intersection point 764 and the intersection point 766, which has been identified as an intersection with an intensity peak corresponding to the sky. After the intensity-discriminating line 780 has been defined, the algorithm proceeds to process element 670 where all pixels above the intensity-discriminating intensity are labelled as sky whilst all pixels above the intensity-discriminating intensity are labelled non-sky. The algorithm then proceeds to the end process element 680. The adapted-watershed image processing algorithm of Figure 6 is just one example of how image processing of the captured image can be performed according to the present technique. In alternative embodiments a greyscale histogram is analysed in different ways from that described above in order to reliably distinguish between sky and ground. For example, the greyscale histogram could be interpolated and searched for minima and maxima to find a "deepest valley" between bumps to determine a threshold between sky and ground. Further alternative implementations are also envisaged

Other aspects and features of the present techniques are described in the following paragraphs.

The present technique relates to a method for sky segmentation using sensor technology and ultraviolet (UV) light, the technique is inspired in part by ant navigation in the natural world.

In a test implementation of the present technique, a standard camera was adapted to allow collection of outdoor images containing light in the visible range, in UV only and in green only. Automatic segmentation of the sky region using UV only, according to the present technique, has been found empirically to be more accurate and more consistent than visible wavelengths over a wide range of locations, times and weather conditions, and embodiments accomplish this with an algorithm that is lower in complexity than previously known techniques. The present technique has been applied to obtain compact binary (sky versus non-sky) images from panoramic UV images taken along a 2km route in an urban environment. Using either sequence Simultaneous Sensor Location and mapping (SLAM) or a visual compass on the captured images has been shown to produce reliable localisation and orientation on a subsequent traversal of the route under different weather conditions.

The recent success of autonomous robotic vehicles [see reference 26] owes much to the availability of specific sensor technologies (LIDAR scanners, IMU, HD cameras, RADAR, GPS) and the continued development of computational resources capable of processing the data they generate. However, for the wider deployment of robots, particularly into price and energy critical markets, it is desirable to bring down the hardware, computational, and energy costs of navigation. With this goal in mind, inspiration can be taken from desert ants, which provide proof of principle that low-power and low-computation methods for navigation in natural environments are possible. Desert ants forage individually and without the use of chemical trails, yet can efficiently relocate their nest or a food source over long distances using visual cues alone [see reference 30]. The ant's compound eye provides very coarse resolution (on the order of 4 degrees per pixel), and thus is unlikely to be capable of extracting and matching many distinct features, a capacity which is fundamental to many current visual SLAM methods (e.g. see reference 5). Rather, the visual system of ants appears fine-tuned to extract a key global feature, namely the panoramic skyline [see reference 16], which behavioral observations have shown is sufficient for guidance in the complex visual habitat of the ants [see reference 7].

Tests implementing the present technique have demonstrated that panoramic skylines (derived from three dimensional sky segmentation process) are sufficient for robot localisation in urban environments [see references 21 , 14, 18]. The skyline (or more precisely, the horizon) has also been investigated as a visual cue for stabilisation in Unmanned Aerial Vehicles [see reference 19]. However, reliable visual segmentation of sky from terrain using conventional visual wavelength imaging appears difficult [see reference 18] particularly in changing weather conditions. Image parsing methods [see e.g. reference 25] can be used to label sky amongst other regions, but the most successful current methods depend on advanced machine learning methods over large numbers of sample images to estimate the characteristics of sky. A range of approaches specific to detecting the sky region only are summarized in reference 20. These include combining colour and texture characteristics, modelling the colour gradient, using connectedness or picture region constraints, and using alternative colour spaces and classification algorithms such as support vector machines. The authors Shen and Wang [see reference 20] have proposed a greyscale method that uses gradient information and energy function maximisation, combined with several picture region assumptions, to obtain results on the order of 95-96% correct classification.

Far-I R (i.e. I R furthest from the visible spectrum) has been used to extract skylines, even after sundown [see reference 14], but is not robust for sky segmentation across variable weather conditions [see reference 4]. Clouds and sky appear very different in near-IR images; indeed I R is often used explicitly for cloud detection [see reference 8].

By contrast, desert ants show sensitivity solely in the UV 350nm and green (510nm wavelengths [see references 17 and 1 1]. The effectiveness of these two wavelengths for sky detection was investigated in reference 16 using a custom sensor with one photodiode sensor for each. The results suggested using UV and green as opponent channels would be the most effective way to segment the sky under dynamic illumination conditions. In a follow up study [see reference 9] using five different wavelengths it was found that UV versus IR performed slightly better. However, because UV-imaging is limited by the filtering in standard camera optics, it has previously received little further investigation in robotics, but according to the present technique, UV imaging has been implemented despite the prejudice in the field to use visible light, I R light or two distinct wavelength bands in conjunction. In reference 3, a UV-receiving camera was paired with an omnidirectional mirror, but used to assess the directional information available in the polarisation of UV-light rather than to detect sky per se (sky segmentation by measuring the degree of polarisation of light is suggested in [see reference 19]). References 23 and 24 describe the use of a UV-passing filter with a camera to enhance the difference between sky and ground regions in images captured such that they include other wavelengths such as visible light, in the context of horizon determination for stabilisation. Neither the accuracy nor consistency of UV versus visible wavelength imaging for detecting sky has been reported in previously-known systems. UV has occasionally been used on robots for purposes other than skyline segmentation, including object marking [see reference 10] and identification [see reference 29].

For proof of concept of the use of primarily UV wavelengths to perform sky segmentation, the utility of UV versus visible light cues for the purposes of sky segmentation according to the present technique has been investigated using an adapted Digital Single Lens Reflex (DSLR) camera system allowing wide spectrum imaging in high resolution. A simple "thresholding" algorithm has been applied to UV-only images and has been shown to be sufficient for accurate (compared to ground truth) and repeatable (comparing time-lapsed images across weather conditions) segmentation. Image capture according to the present technique has been used to create panoramic segmented sky images that support navigation in an urban environment. At least some key features of embodiments may be rapidly realised using off-the-shelf components (e.g. CCD without colour filters) providing a new class of low-cost outdoor navigation system.

The aims of the empirical investigation of the present technique study were: firstly, to assess the relative accuracy of sky segmentation using different wavelengths, over a range of different locations and conditions; secondly to assess the consistency of segmentation for UV versus visible wavelengths in the same location but at different times of day and weather conditions; and third to test the effectiveness of the segmented sky shape as a cue for robot navigation. The imaging device used for the empirical tests was a device like the embodiment of Figure 2, the data sets collected for each test, the segmentation and navigation methods used will now be described. Considering full spectrum imaging, sample images in both UV and visible spectra have been taken by replacing the internal cut filter of a standard DSLR camera (e.g. Nikon [RTM] D600) by a fused quartz window, and fitted with a high UV transmission wide angle lens (Novoflex Noflexar [RTM] 3.5/35mm). Images could then be recorded with different filters: Thor Labs FGB37 bandpass filter (335 - 610 nm) to sample only visible light; Baader U-Filter (350 nm, half-width 50 nm) to approximate the ant's UV photoreceptor; and Knight Optical [RTM] 520FCS5050 bandpass filter (515nm) to approximate the ant's green photoreceptor profile. The camera aperture was set to f/16, sensor sensitivity (ISO) to 200 and for each image an appropriate shutter speed was automatically selected to provide proper exposure given the chosen aperture. All images were saved using the uncompressed 14 bit NEF file format (a proprietary Nikon [RTM] raw image format) and subsequently de-mosaiced to a true colour 16 bit Tagged Image File format (TI FF). Visible light images were stored in RGB format, the CMOS red channel only used for UV images, and a greyscale conversion of all three channels for green filtered images. To mimic the bi- chromaticity of the ant visual system, UV-G images were created by combining separate UV and Green intensity images into a two dimensional false colour space. All images were then scaled down by a factor of ten along each dimension to give a resolution of 604x403 pixels prior to analysis.

Around seven hundred and fifty photographs were taken in a variety of locations, including the natural habitat of desert ants (Seville, Spain) and wood ants (Sussex, UK) and a series of urban settings. UV-passing, green-passing and visible light-passing filters were used in turn. A subset of eighteen images was selected representing a wide diversity of environments and conditions (see table I for example images) for which ground-truth sky segmentation was established. These included images expected to be difficult to segment due to surfaces reflecting the sky, such as water and windows. Semi-automated labelling of sky and ground pixels was used through iterative applications of a "GrowCut" algorithm with user-correction [see reference 27]. It was not possible to use pre-existing "benchmark" labelled datasets as these did not include UV information.

Table I of Figure 9 shows the accuracy of sky segmentation across different locations. A subset of our hand labelled image database is shown in Figure 9 to illustrate the diversity of images collected. Image triplets (Visible, UV-G and UV) were sampled in locations chosen to assess the performance of sky segmentation in diverse surroundings. The percentage of correctly labelled pixels, when compared to a user defined ground truth, is shown for each image set. Mean scores (final row) are for the full database. Overall, clustering performance is found to be best using k-means on UV images, however similar results can be achieved with the computationally simpler adapted-watershed algorithm. Timelapse Database

A time-lapsed image database was gathered from a single vantage point by mounting the camera on a secure platform and taking images across four days at ten minute intervals in varying weather conditions (15:35-17:55 low light, 10:25-1 1 :55 overcast, 10:50-1 1 :00 sunny and 10:50-1 1 : 10 partial cloud). As before, all three filter types were used at each time point.

Urban Route Database

To allow the entire panorama to be recorded in a single image, the full spectrum camera was mounted above a panoramic mirror (Kugler [RTM], GmbH), fixed via a UV-passing quartz cylinder with 2mm wall (see Fig. 8a), and images were recorded using the UV- passing filter. A Sony [RTM] Bloggie MHS-PM5, fitted with Bloggie [RTM] 360 video lens kit attachment, was used to simultaneously capture panoramic images in the visible light spectrum. The combined system was mounted on a pole to elevate the horizon above a human carrier, with an approximate height of recording of 200 cm. Both cameras recorded video while the carrier walked a 2 km route that included narrow streets, city parks and squares, with moving traffic and pedestrians. True position was logged once per second using a GPS watch (Suunto Ambit [RTM]). For testing navigation two videos of the route were used that were sampled on different days, at different times of year (January and April, introducing a change in the foliage), under different weather conditions (overcast and blue sky with clouds), and with different sun positions (1 1 :40 and 16:47). Each video was converted to "JPG" (Joint Photographic Experts Group) images at 1 frame per second, providing two sets, consisting of 1337 and 1452 images respectively.

To remove any distortion effects caused by camera movement when changing filters, images were aligned prior to segmentation by mutual information based image registration [see references 13 and 1]. K-means (K=2, r=7, sq. Euclidean distance) and GMM clustering algorithms were implemented to assess both the accuracy and computational ease with which the sky could be automatically segmented from terrain. Pixels were clustered using image intensity only, ignoring any spatial connectivity information. These methods allowed comparison of performance despite the three image types having differing dimensions: UV=1 D; UV-G=2D; and visible=3D (RGB).

A simpler segmentation algorithm tailored to the 1 D UV intensity images was also tested. A histogram is computed from the intensity image, excluding the darkest and brightest two percent of pixels, and the adapted-watershed algorithm described in references 28 and 31 was applied until only two segments of the histogram remained (see Figures 6 and 7 and their associated descriptions above). An intensity value midway between the closest points of the two histogram segments (two main peaks) was then set as the threshold for ground/sky segmentation. This method is robust to multiple maxima caused by a visible sun and ground reflectance, which can be problematic for alternative algorithms in previously known systems.

Navigation Tests

OpenSeqSLAM [see reference 22], an open source Matlab [RTM] implementation of SeqSLAM [see reference 15], was used to test whether panoramic UV-segmented images contain sufficient information for localisation. Images were preprocessed by applying a mask to disregard sectors of the image that were not parts of the panorama in the parabolic mirror. The simple adapted watershed segmentation algorithm described above was then applied to produce 90x90 binary images (sky versus non-sky). The sky region was isolated by boundary tracing from the centre of the image, and shifted by centralising the centre of gravity to reduce effects of tilt (Fig. 8b). Performance of the UV-segmented image set was compared to regular SeqSLAM using a set of visible light panoramic images recorded at the same locations and with a similar number of bits per image (16x16, 8 bit greyscale).

Localisation on the April route was tested with the January reference images and vice versa. To distribute effects caused by the two sequences being particularly in or out of sync, a reference set was compared starting at the first frame of the route and sampled every 10 seconds against 10 test sets, starting with the first, second, third etc. frame of the other route and sampled every 10 seconds thereafter. SeqSLAM was run in both directions for each of the resulting nineteen route pairs. SeqSLAM first performs an image difference function on the current image and all training images, stores the results in a matrix, and uses contrast enhancement to find the local best matches. A familiar sequence is then found by scoring trajectories across the matrix to look for those with the combined best match. The trajectory with the lowest score was considered a match if it was significantly smaller than all other scores outside a sliding window (R_window=10).

An alternative navigation method that has been suggested as a possible mechanism used by ants, i.e. the visual compass algorithm [see reference 6, 12] was also tested. The basic assumption of this algorithm is that rotating a test image until the difference from a training image is minimised will indicate an approximately correct heading, under the assumption that captured training images are generally oriented in the direction of forward travel along the route. Although some versions of this algorithm require knowledge of the sequence to constrain which images to compare [see reference 12] it has also been shown to work effectively by taking the minimum over all rotations over all training images [see reference 2], i.e. , assuming no sequence memory. This implicitly also returns a location estimate: the training image most similar to the test image after rotation.

Using the extracted skyline image sets (obtained from sky segmentation) at full resolution (691x691 pixels) a sub-sample of one image every 10 seconds from the test set was compared at 90 4 degree rotations against every image in the full 1 frame per second training set, by summing the output of an XOR function. The location and rotation producing the smallest overall difference was chosen as best match. A comparison was made to both greyscale and colour visible light images, unwrapped and scaled to a similar number of bits; 180x28 and 90x14 pixels respectively.

Automatic Segmentation of Sky is Best Using UV Intensity Images

Table I in Figure 9 shows the results of automatic sky segmentation, compared to manually labelled ground-truth, in the first image-database. Twelve example images are shown; the mean accuracy is based on eighteen images. Segmentation using K-means clustering is most accurate using UV intensity images, with a mean accuracy of 96.8%. This is found to be better than visible light (mean accuracy 89.9%, paired Wilcox test p=0.0018) and UV-G (mean accuracy 91 .5%, paired Wilcox text p=0.025). The accuracy for UV is also found to be less variable (standard deviation 3.6) than either visible light (standard deviation 12.3, F-ratio test p< 0001 ) or UV-G (s.d. 9.5, F-ratio test p=0.0002). Segmentation in the visible spectrum was not improved using the more robust Gaussian mixture model classifier (mean accuracy 86.0%). Visible light sky segmentation was also tested in other colour spaces, such as the A and B dimensions of LAB space, however no notable improvements were observed over RGB for any of our clustering algorithms. The simple adapted watershed algorithm for UV performs just as well as K-means clustering (mean accuracy 96.9%) but at less computational cost (O(N) instead of O(kDN)), where k is the number of clusters, D the dimensionality of our pixel space and N the number of pixels). The level of accuracy obtained is comparable with that reported in other state of the art sky segmentation methods, e.g. 96.05% reported in [see reference 20]. In fact a proportion of the remaining error was simply due to vignetting of the overlapping lens hood (see Fig. 2 middle row). We also note that the ^'ground truth' labelling was performed on the visible images and so was if anything biased against UV.

Figure 10 shows the underlying reasons for better performance of UV. Figure 10 shows a typical example of conditions in which UV-intensity segmentation outperforms both visible and UV-G images. Performance is worst using visible light as sky and ground are not easily separable when clouds are present. In contrast, sky and ground naturally separate along the UV dimension giving robust segmentation in UV and UV-G images. Raw images are shown in the left column. The central column shows the spread of image pixels plotted in their respective colour space, coloured with respect to the attempted automatic k-means segmentation. In each scatter plot a random subsample of 200 pixel values was used. The right column shows the resulting segmented images with white representing sky and black representing ground. Note that the black marks in the corners of the UV pictures are the result of vignetting due to an overlapping lens hood that is part of the Baader U-filter. In the visible light image of Figure 10 (top left), the bright white clouds form a cluster. As the classifier is constrained to two groups, the blue sky is incorrectly labelled the same as the ground. This is the most common cause of error when using visible light, but reflectance from buildings and sun flare also cause segmentation irregularities. In contrast, in the UV-G image the clouds are labelled with the sky, and only reflection from the buildings causes some segmentation issues. The pixel-plot shows that the successful separation is attributable mostly to the UV channel, and in fact the segmentation is improved by using the UV channel only, which does not classify strong lighting on the buildings as sky.

The second image-database allows us to test both the accuracy (versus ground truth), and the reliability (versus images from the same location) of sky segmentation across weather and lighting conditions. This is important as it may be acceptable for a navigating robot to misclassify some portion of the ground as sky (or vice versa) provided the classification remains consistent over time. Accuracy was compared against user labelled ground truth, and here again it has been found that segmentation is more accurate using UV intensity images (mean=99.5%, standard deviation (s.d.)=0.084 ) than visible (mean=95.2%, s.d.=5.327 ). Reliability was assessed by measuring the mean entropy and Pearson correlation coefficients of the binary labelled images. Specifically, the mean entropy per pixel location, in labelled UV and visible images was calculated as

(1)

where x_it is the intensity value of a pixel at location /^' at time t, and T is the number of images in the set. The Pearson correlation is given by

where v"{ , )indicates the sample covariance between two images x and y and var'{ } the sample variance of an image. Figure 1 1 presents the resultant entropy and correlation plots for visible and UV image sets. It is clear from Figure 1 1 that pixels in the visible domain are highly variable (mean entropy=0.1877) particularly in the crucial area where buildings meet the sky. Reliability of sky segmentation across weather conditions is demonstrated by Figures 1 1 a to c. Figure 1 1 a shows examples of time-lapse images sampled under differing weather condition (note that the UV pixel intensities have been increased 5* to make image more visible, but are still extremely low in the third example). Figure 1 1 b shows pixel-wise entropy and Figure 1 1 c shows Pearson correlation coefficients computed across the image dataset respectively, both indicating high variance in visible images but consistent labelling using UV images.

There is also interference from sunlight on buildings and street lights. This variance results in low correlation coefficients between images throughout the data-set. Labelled images that correlate poorly in the heat map in Fig. 1 1 c (images 5, 15 and 20) corresponded to the 3 images shown in the left column of the Fig. 1 1 a, where cloud, sunlit buildings and low lighting caused discrepancies in segmentation. The causes of these problems are clearly visualised by viewing the plot showing average entropy per pixel (Fig. 1 1 b). In contrast, UV images have a low mean entropy per pixel (0.0048), producing a highly correlated and thus reliable dataset across conditions, even at low light levels.

UV Sky Segmentation Images Are Sufficient For Navigating Cities

A set of images recorded along a route can be used in several ways for subsequent navigation. We first tested whether binary sky/non-sky images derived from panoramic UV images recorded along a 2km urban route could be used for localisation (or loop-closing) using SeqSLAM as described above. Overall, using a window length of ten images, we can obtain a matching precision of 100% (no false positives) with a recall of 81 %; at 100% recall, the precision is 94%; these values compare well to previous results for this algorithm [see reference 15]. Figure 12 shows where matching is unsuccessful, which appears to correspond to locations under trees, where the sky is either obscured or its segmented shape changing too rapidly for the sparse temporal sampling used here. The result for binary UV images is only a few percent below that obtained using standard SeqSLAM on the equivalent panoramic greyscale visible light images (precision 97%), which deals better with situations where there is little or rapidly changing sky shape information (under trees) as it can still exploit information in the ground region. On the other hand, standard SeqSLAM tended to fail in areas where the ground region is relatively featureless but there are substantial changes in the sky region (clouds) between training and test sets. In Figure 12 the use of sky-segmented images to navigate a city is shown. Figures 12 (a) and (b) plot the example results of the SeqSLAM localisation trial taken from the worst test set for UV (left) and the worst for visible light (right). Some of the dots represent locations that were correctly recognised and some of the dots represent locations that were not. In the top right corner of these figures the estimated location index of the training set is plotted against the location indices of this test set. In the bottom left corner an example panoramic image is shown, corresponding to an incorrectly localised image in this test set. The branches and clouds were typical features that can sometimes cause the algorithm to fail. Figures 12 (c) and (d) show the heading direction that the visual compass algorithm would select in order to retrace the route (left for UV, right for visible light), with many headings corresponding to a correctly matched location being found. The Map data came from Google [RTM], Infoterra Ltd [RTM] & Bluesky [RTM].

While SeqSLAM is highly effective for localisation if following the same route, it may not provide a mechanism to stay on the route. Hence the visual compass algorithm was also tested, as described above which produces both a location match and an estimate of the heading direction needed to stay on route. The obtained precision using binary UV images was 84.7%. Figure 12c shows the heading direction suggested by the visual compass at each point on the route; it can be seen that these align well with the required direction to stay on the route. Using the equivalent visible light images produced a significantly worse performance with a precision of only 44.1 % and many heading direction choices that would lead away from the route (Figure 12d).

In summary, it has been suggested that UV light sensing in desert ants is a specialisation that supports their impressive navigation by enhancing sky-ground contrast. Experiments using embodiments of the image capturing apparatus according to the present technique have demonstrated here that UV images are highly effective for sky segmentation, that the segmented sky shape is robust to changing light and weather conditions, and that the resulting binary sky/non-sky images can support effective navigation. For a wide range of locations and light conditions, automated sky segmentation has been demonstrated to be more effective for UV than visible light. Although only a limited set of images were hand-segmented for ground truth evaluation, visual inspection of UV segmentation in a larger collection of images suggests broad reliability, in particular, robustness against all kinds of cloud cover. The uni-dimensional signal also allows the application of a computationally cheap segmentation algorithm that is just as effective as K-means clustering.

There were a few cases where ground objects, such as particular kinds of metal roofing, were found to have had high UV reflectivity. However, for navigation, it is less important to find the skyline (or sky/ground threshold in three dimensions) exactly than it is to find it consistently. By testing sky-ground segmentation for images taken from the same location on different days, at different times of day, and under different weather conditions, it has been found that UV images have an extremely high consistency in pixel classification, around forty times better than visible light.

The navigation potential of a panoramic UV camera system was demonstrated by recording 2km routes in an urban environment on two different days with different weather conditions. Using either SeqSLAM or a visual compass, matching of UV-segmented binary images of the sky contour from one day against another supported successful localisation. For the visual compass a reliable indication of the correct heading direction to continue along the route was also obtained. Using non-segmented visible light images of a similar file size, SeqSLAM was still effective for localisation, but determining location or heading direction from the visual compass was far less effective. Failure of the visual compass is generally due to mismatch or aliasing, i.e. , a training image far from the current location provides, at some arbitrary rotation, a closer match than the nearest image. This suggests its effectiveness may be improved (and its application sped up) by using SeqSLAM as a pre-processing step to find the best match, then finding the minimum over image rotations for this image only.

The skyline (in two or three dimensions) is a particularly consistent aspect of urban (and some other) environments. Many changing features such as traffic and pedestrians fall entirely below the skyline, and different lighting conditions such as sun falling on one or the other side of street make little difference to the segmented shape. Weather changes occur above the skyline. Consequently it appears that a substantial amount of information from the panoramic image is retained by reducing images to binary sky versus non-sky: SeqSLAM navigation performs almost as well on such images sky segmented images as on full grey scale images. The few situations in which the system failed in the particular test performed were in locations where very little sky was visible (e.g., under scaffolding, or under trees). The considerable difference in foliage between our two datasets, and the ten second time intervals between frames could cause a notable difference in sky shape directly above the camera between training and test sets in this situation. This could be rectified by smoothing over recent frames to filter out fast moving shapes. In alternative embodiments of the present technique, rather than using binary images, the UV threshold could be used to mask out the sky region (including uninformative features such as clouds) but information below the skyline (considering two or three dimensions) , possibly including light from visible light channels, could be retained (rather than filtered out before the CCD image is captured) to improve robustness.

There are a number of potential applications for the image capture system according to the described embodiments. It could be used effectively in urban environments where GPS can be unreliable. The nature of the segmented image also makes it convenient and efficient to perform corrections to tilt by simply offsetting the image depending on the centre of gravity of the sky pixels. It could therefore also be potentially useful for robots in rugged environments, where SLAM needs to be reliably carried out on uneven terrain. Sensitivity to UV light at low light levels could be improved by using a CCD chip without colour filters, rather than the converted visible light camera used in the empirical tests described above. The relatively simple computation to segment sky could be implemented in dedicated hardware such as a System on Chip, providing a compact and low-power sky detection sensor, which could offer cheap high precision localisation in many outdoor autonomous robot applications.

It will be appreciated that embodiments of the present invention can be realized in the form of hardware, software or a combination of hardware and software. The hardware can be analogue or digital electronic circuits, optical components and image capture components.. The software may be configured to control processing of the captured image. Any such software may be stored in the form of volatile or non-volatile storage, for example a storage device like a ROM, whether erasable or rewritable or not, or in the form of memory, for example RAM, memory chips, device or integrated circuits or on an optically or magnetically readable medium, for example a CD, DVD, magnetic disk or magnetic tape or the like. It will be appreciated that the storage devices and storage media are embodiments of machine-readable storage that are suitable for storing a program or programs comprising instructions that, when executed, implement embodiments of the present invention.

Accordingly, embodiments provide a program comprising code for implementing apparatus or a method as claimed in any one of the claims of this specification and a machine- readable storage storing such a program. Still further, such programs may be conveyed electronically via any medium, for example a communication signal carried over a wired or wireless connection and embodiments suitably encompass the same.

List of References

[1 ] Artyushkova, K. Automatic Image Registration using (Normalized) Mutual Information for users of I P toolbox, 2006. URL http://www. R-project.org/.

[2] B. Baddeley, P. Graham, P. Husbands, and A. Philip- pides. A model of ant route navigation driven by scene familiarity. PLoS Computational Biology, 8(1 ):e1002336, January 2012. doi: 10.1371/journal.pcbi.1002336.

[3] N. Carey and W. Sturzl. An insect-inspired omni- directional vision system including UV-sensitivity and polarisation. In Computer Vision Workshops (ICCV Workshops), 201 1 I EEE International Conference on, pages 312-319. I EEE, 201 1. doi: 10.1 109/ICCVW.201 1. 6130258.

[4] L. Chapman, J. E. Thornes, J-P. Muller, and S. McMul- droch. Potential applications of thermal fisheye imagery in urban environments. Geoscience and Remote Sensing Letters, IEEE, 4(1 ):56-59, 2007. doi: 10.1 109/LGRS. 2006.885890.

[5] M. Cummins and P. Newman. FAB-MAP: Probabilistic Localization and Mapping in the Space of Appearance. The International Journal of Robotics Research, 27(6): 647-665, June 2008. doi: 10.1 177/0278364908090961 . [6] M.O. Franz, B. Scholkopf, H.A. Mallot, and H. H. Bulthoff. Where did I take that snapshot? Scene-based homing by image matching. Biological Cybernetics, 79 (3): 191-202, 1998. doi: 10.1007/s004220050470.

[7] P. Graham and K. Cheng. Ants use the panoramic skyline as a visual cue during navigation. Current Biology, 19 (20):R935-7, November 2009. doi: 10.1016/j.cub.2009.

08.015. [8] D. I. Klebe, R. D. Blatherwick, and V.R. Morris. Ground- based all-sky mid-infrared and visible imagery for pur- poses of characterizing cloud properties. Atmospheric Measurement Techniques, 7(2):637-645, 2014. doi: 10.5194/amt-7-637-2014.

[9] T. Kollmeier, F. Roben, W. Schenck, and R. Moller. Spectral contrasts for landmark navigation. Journal of the Optical Society of America. A, Optics, Image Science, and Vision, 24(1 ): 1— 10, January 2007. doi: 10.1364/JOSAA.24.000001.

[10] S. Komai, T. Kuroda, and M. Takano. Development of Invisible Mark and Its Application to a Home Robot. In Service Robotics and Mechatronics, pages 171-176.

2010. doi: 10.1007/978-1-84882-694-6_30.

[1 1 ] T. Labhart. The electrophysiology of photoreceptors in different eye regions of the desert ant, Cataglyphis bicolor. Journal of Comparative Physiology A, 158(1):

1-7, 1986. doi: 10.1007/BF00614514.

[12] F. Labrosse. The Visual Compass : Performance and Limitations of an Appearance- Based Method. Journal of Field Robotics, 23(10):913-941 , 2006. doi: 10.1002/rob.

[13] F. Maes, A. Collignon, D. Vandermeulen, G. Marchal, and P. Suetens. Multimodality image registration by maximization of mutual information. I EEE transactions on medical imaging, 16(2): 187-98, April 1997. doi: 10.1 109/42.563664.

[14] J. Meguro, T. Murata, Y. Amano, T. Hasizume, and J. Takiguchi. Development of a Positioning Tech- nique for an Urban Area Using Omnidirectional In- frared Camera and Aerial Survey Data. Advanced Robotics, 22(6-7): 731-747, January 2008. doi: 10.1 163/156855308X305290.

[15] Michael J Milford and Gordon Fraser Wyeth. Seqslam: Visual route-based navigation for sunny summer days and stormy winter nights. In Robotics and Automation (ICRA), 2012 I EEE International Conference on, pages1643-1649. I EEE, 2012.

[16] R. Moller. Insects could exploit UV-green contrast for landmark navigation. Journal of theoretical biology, 214 (4):619-31 , February 2002. doi: 10.1006/jtbi.2001 .2484.

[17] M. I . Mote and R. Wehner. Functional characteristics of photoreceptors in the compound eye and ocellus of the desert ant, Cataglyphis bicolor. Journal of comparative physiology, 137(1 ):63-71 , 1980. doi: 10.1007/ BF00656918.

[18] S. Ramalingam, S. Bouaziz, P. Sturm, and M. Brand. Geolocalization using skylines from omni-images. In Computer Vision Workshops (ICCV Workshops), 2009 I EEE 12th International Conference on, pages 23-30. I EEE, 2009. doi:

10.1 109/ICCVW.2009.5457723. [19] A.E. R. Shabayek, C. Demonceaux, O. Morel, and D. Fofi. Vision Based UAV Attitude Estimation: Progress and Insights. Journal of Intelligent & Robotic Systems, 65 (1 - 4):295-308, 2012. doi: 10.1007/s10846-01 1 -9588-y.

[20] Y. Shen and Q. Wang. Sky Region Detection in a Single Image for Autonomous Ground Robot Navigation. International Journal of Advanced Robotic Systems, 10: 362, 2013. doi: 10.5772/56884.

[21 ] F. Stein and G. Medioni. Map-based localization using the panoramic horizon. I EEE Transactions on Robotics and Automation, 1 1 (6): 892-896, 1995. doi: 10.1 109/70.478436.

[22] N. Sunderhauf, P. Neubert, and P. Protzel. Are We There Yet? Challenging SeqSLAM on a 3000 km Journey Across All Four Seasons. I n Proc. of Workshop on Long-Term Autonomy, I EEE International Conference on Robotics and Automation (ICRA), Karlsruhe, Germany. ,2013.

[23] M. H. Tehrani, M.A. Garratt, and S. Anavatti. Gyroscope offset estimation using panoramic vision-based attitude estimation and Extended Kalman Filter. In CCCA12, pages 1-5. I EEE, December 2012. doi: 10.1 109/CCCA. 2012.6417863.

[24] Mohsen H. Tehrani, M. A. Garratt, and S. Anavatti. Horizon-based attitude estimation from a panoramic vi- sion sensor. In Ghose Debasish, editor, IFAC-EGNCA

2012, pages 185-188, February 2012. doi: 10.3182/20120213-3-I N-4034.00035.

[25] J. Tighe and S. Lazebnik. Superparsing. International Journal of Computer Vision, 101 (2):329-349, October 2012. doi: 10.1007/s1 1263-012-0574-z.

[26] Urmson, C. The self-driving car logs more miles on new wheels, August 2012. URL http://googleblog.blogspot.hu/2012/08/ the-self-driving-car-logs- more-miles-on.html.

[27] V. Vezhnevets and V. Konouchine. GrowCut: Interactive multi-label ND image segmentation by cellular automata. In Proc. of Graphicon, pages 150-156, 2005.

[28] H. Wajima, T. Makino, Y. Takahama, T. Sugiyama, and A. Keiichi. An Autonomous Inspection Robot System for Runways Using an Ultraviolet Image Sensor. In Intelligent Autonomous Systems 7, pages 357-364, 2002.

[29] R. Wehner. The architecture of the desert ant's navi- gational toolkit (Hymenoptera: Formicidae). Myrmecol News, 12:85-96, 2008.

[30] N.S. Zghal and D.S. Masmoudi. Improving watershed algorithm with a histogram driven methodology and implementation of the system on a virtex 5 platform. International Journal of Computer Applications, 9(1 1 ): 29-35, 2010. doi: 10.5120/1435-1934.

Claims

CLAIMS: 1. An imaging apparatus configured to discriminate between sky and non-sky, the apparatus comprising:

a UV-passing image forming element;

a UV-sensing image capturing sensor for capturing an image projected from the UV-passing image forming element, the image comprising at least one of sky and non-sky image regions;

an image processing unit for performing discrimination between sky and non-sky regions of the captured image using primarily UV wavelengths.

2. An apparatus as claimed in claim 1 , wherein the image processing unit is configured to perform an intensity categorisation of the image regions of the captured image, the intensity categorisation being adaptive to image characteristics comprising at least one of: light levels; image capture time; weather conditions; and illumination angle.

3. An apparatus according to claim 2, wherein the intensity categorisation performed by the image processing unit uses a single-dimension intensity metric comprising primarily UV wavelengths.

4. An apparatus as claimed in any one of the preceding claims wherein the image processing unit is configured to performing the discrimination between sky and non-sky image regions using a subset of primarily UV wavelengths selected from the UV wavelength range of 200nm to 400nm.

5. An apparatus as claimed in any one of the preceding claims, wherein the UV- passing image forming element comprises one of a substantially hemispherical lens and a fish-eye lens.

6. An apparatus as claimed in any one of the preceding claims, wherein the UV- passing image forming element comprises at least one curved mirror.

7. An apparatus as claimed in any one of claims 1 to 4, wherein the UV-passing image forming element comprises an array of image forming elements arranged on a curved surface, each image forming element having a different field of view such that together the array of image forming elements is configured to enable substantially simultaneous capture of a sky panorama.

8. An apparatus as claimed in claim 7, wherein each of the image-forming elements of the array of image-forming elements has at least one of: a respective UV-sensing image capturing element and a respective UV-passing filter forming an integral unit with the respective image-forming element mounted on the curved surface.

9. An apparatus as claimed in any of claims 1 to 4, wherein the UV-passing image forming element and the UV-capturing image forming element are disposed together as an image- capturing unit on a rotatable mounting such that a sky panorama is obtained by rotating the image-capturing unit to obtain a composite image of a solar hemisphere.

10. An apparatus according to any one of the preceding claims, wherein the UV- sensing image capturing element is configured to be sensitive primarily to UV light.

11. An apparatus according to any one of the preceding claims, wherein the image processing unit is configured to receive greyscale image data from the image-capturing sensor and to bin the greyscale image data according to pixel intensity values to obtain binary image data, the binary values discriminating between pixels associated with sky and pixels associated with non-sky.

12. An apparatus according to claim 1 1 , wherein the image processing unit is configured to establish a threshold intensity value from the greyscale image data such that pixels having an intensity value below the threshold are identified as non-sky pixels and pixels having an intensity value above the threshold are identified as sky pixels.

13. An apparatus as claimed in claim 12, wherein the threshold intensity value is determined by identifying in a histogram of the binary image data an intensity peak associated with sky pixels and an intensity peak associated with non-sky pixels and selecting the threshold intensity value to be an intensity located between the sky peak and the non-sky peak.

14. An apparatus as claimed in any one of claims 11 to 13, wherein a centre of gravity of pixels corresponding to sky in the binary image data is used to perform tilt correction in a navigation system.

15. A robot navigation system comprising the apparatus of any one of claims 1 to 14.

16. A method of image capture comprising the steps of:

directing light from a skyline panorama through a UV-passing image-forming element;

capturing an image projected from the UV-passing image forming element using an image-capture sensor, the image comprising at least one of sky and non-sky image regions;

performing discrimination between sky and non-sky image regions of the captured image by performing processing of pixels of the captured image using primarily UV wavelength components of the captured image.

17. An apparatus substantially as hereinbefore described with reference to Figures 1 to 5 and 8a of the accompanying drawings.

18. A method substantially as hereinbefore described with reference to the accompanying Figures 6 and 7 of the accompanying drawings.