DE3033785A1 - Image processing target acquisition and tracking - uses Bayes classifier to classify target and terrain and controls measurement data acquisition - Google Patents

Abstract

An arrangement for detecting and tracking a target contains a sensor which responds to electromagnetic radiation. It automatically optimises the distinction between the target and the terrain and is very simple. The measurement values produced by the sensor are reduced to two intensity levels for the target and the terrain. The target and terrain are classified in a classifier using Bayes Theorem. The classifier controls the acquisition of measurement data via a switching function. The classifier a posteriori signal is fed to a coordinate analyser which determines the target coordinates using a maximum target point probability method. A delay element forms a signal which acts on the a priori signal from the posteriori signal. The target coordinates are fitted with time and a difference signal between the predicted and measured maxima fed to the delay element.

Description

Device for Detecting and Tracking a Target The invention deals with a device for detecting and tracking a target, which has a sensor responsive to electrical radiation.

Establishments of the aforementioned type are from an extensive range known in the art (see, for example, DE-AS 26 02 838, DE-AS 25 47 798, DE-AS 24 41 640). The image data processing described there can very generally lead to a measured value processing problem be traced back, with target tracking being a special case of such Measurement processing is. Essentially, target tracking is based on distinction of two processes, namely target and terrain, so that those used for target tracking Facilities consist of three parts, the processing of measured values for data reduction, the classification of the measured values into the target and terrain classes and the determination the destination coordinates from the classified measured values.

The previously known facilities either do not meet the in They set expectations because they make the necessary distinction between goal and terrain not with sufficient accuracy or not at sufficient speed deliver, or they are too expensive so that they can no longer be realized or in institutions, like missiles or similar, because of their weight and space requirements cannot be used.

The object of the invention is to provide a device of the type mentioned at the beginning Specify type, with the help of which an automatic optimal classification between Objective and terrain can be carried out and in particular by simple Realization excels.

This object is achieved by a) that the supplied by the sensor Measured values recorded in one measured value acquisition and on two brightness densities for target and terrain are reduced, b) that these are used as input variables to a classifier are supplied, the classifier using the Bayes theorem the classification in the target area and in the area and controls the measurement of the measured value via a switching function, c) that the a posteriori signal of the classifier is both a target co-ordinator is supplied, which the target coordinates based on a center of gravity, which is applied to the probability of presence of the target points, d) as well as a delay element that is derived from the posteriori signal on the a-priori information produces a retroactive signal e) and that a filter the Filters target coordinates over time and the delay element a difference signal between the predicted target center of gravity and the measured center of gravity.

The device is particularly characterized by the fact that as a measured value or only the brightness in one point is used as a feature for target tracking, i.e. positional relationships between measuring points are not taken into account, so that each measuring point can be processed independently of its neighbors. this leads to to a very strong simplification of the classifier, so that one can without much effort the Processing can extend to all measuring points, including the background information, which makes up the largest part of the measured values can be included. For this only the requirement has to be made that the background is stationary and is also mapped in this way in the measurement matrix, i.e. that the position of the sensor in relation to is known or stabilized on a fixed reference, otherwise the a priori information, that the background is stationary and cannot be used.

The facility's inherent recursion leads to a desired one Learning effect in such a way that the estimation of the position of the target increases with the number of measured values converges to the true value. The testing of the invention has showed that with the help of the Baves rule an optimal classification in the statistical sense can be carried out.

The invention is explained in more detail with reference to the single figure.

Before they are explained in more detail, a few general Comments on image processing are made.

As is known, the only information that you can get from one is based on Electromagnetic radiation responsive sensor receives, the brightness, possibly. also the color, which is measured e.g. in a rectangle of 256 x 256 pixels and their location in the eating matrix.

In addition, however, one knows and this is for further processing it is necessary that the terrain is stationary, the target compact and mostly fluid, i.e. the measured values of the target have no defined positional relationship to one another, and besides the goal obeys a movement model. This means that target tracking can be divided into two sections be structured, namely the measurement of the information and the processing of the Information, with additional information that is not measurable but can be derived from physics and, to a decisive extent, from the one to be solved Task determined to play an essential role in processing.

As the figure now shows, the device consists essentially of a measured value processing 10, in which the from a sensor, not shown Delivered measured values are detected and reduced from a classifier 11, the based on the Bayes' theorem with maximum probability the measured values due to correctly classified based on his a priori knowledge and that of a switching function U controls the measured value processing 10, and controls the input variables f1, f2 of the classifier 11 get from the measured value processing 10, a target coordinate determiner 12, with which the target coordinates are determined and its input signal q from the classifier 11 receives, the signal q also being fed to a delay element 13, in that from the previously calculated a posteriori probabilities of the eliminator 11 new a priori probabilities are calculated and the classifier 11 are supplied as signal v. There is also a 1less filter with which the Position of the target can be accurately estimated and at the exit of which the position of the target can be removed in x and y coordinates. The filter also gives a signal Ai, Aj ab, which is fed to the delay element 13.

Overall, it can be said that the a priori and a posteriori information v or q and the input variables of the classifier fl, f2 are understood as state variables so that it is necessary when initializing the device to have or estimate certain initial values for this state variable. The classifier 11 itself is optimized in such a way that it is based on measured values with maximum probability correctly classified his a priori knowledge. For this purpose, the well-known Bayes' theorem used.

From the literature (see Günther Meyer Brötz, Jürgen Schürmann "Methods the automatic time recognition "R. Oldenburg-Verlag, Munich-Vienna 1970) known that Bayes' theorem provides an optimal decision regarding a Cost function, where the mean cost risk becomes a minimum and all possible Wrong decisions are charged with the same costs. For the invention uses Bayes' theorem to mean that from the two competing processes Target and terrain the decision should be made whether a measured pixel belongs to the target or terrain.

This rule can be formulated as follows.

If pjfj> pjfj # j # j, then process J was measured.

The following definitions are made.

pj = "a priori" probability for the occurrence of the process j fj = the probability density of x that x is the measured brightness of the process j x: measured value of the normalized brightness (intensity) at a point with 0 # x 1 j: running index for the possible processes i # j # n 3: number of the process that is decided on according to the classification rule 1 # J # nn: number of possible processes (here: n = 2: target / terrain Based on the equation, one can also ea- posteriori "- probabilities for the measurability of a process i. This probability is called w. and the following applies: wj ~ pj. fj 1 1 i wi = = C Pilz In addition, the following applies to probabilities and thus for implementation in classifier 11 The decision equation for two processes n = 2 is: J = 1 for w1> w2 2 = 2 for w2> w1 It is unknown how Pi and w are determined and what their meaning is. This is explained below.

When tracking the target, the process can split up the terrain again are in the foreground and background. An object can be behind a foreground be present without it being visible. For this hidden presence one can also specify a probability, the so-called presence probability. This contains valuable information, since you can also find coordinates when the object is covered can specify for his position. However, the foreground can only be recognized when the object disappears behind it, i.e. the foreground only appears with the Presence of the object visible or cleared up. In this case, despite being high "a priori" visibility probability (measurability probability P1) only a low "a posteriori" visibility probability w1 is calculated. That stirs hence the fact that the brightness x is not related to the object density f1, but to the terrain density f2 heard.

Since an object cannot disappear from the scene, so it cannot more is present, one can have a certain probability of being present of the foreground indicate. For the mathematical derivation of this Relationships are used to denote the individual probabilities The following agreements are made, which are also referred to again and again in the following is taken.

pO a-priori probability of visibility for foreground Pl a-priori "for object P2 a-priori" for terrain (foreground and background where a-posteriori probability of visibility for foreground w1 a-posteriori "for object w2 a-posteriori" for terrain v0 a-priori Presence probability for foreground v1 a-priori "for object v2 a-priori 8 for terrain q0 a-posteriori probability of presence for foreground q1 a-posteriori "for object q2 a-posteriori II for terrain It is not all above defined sizes are required, but they are listed for the sake of completeness.

One can use the following equations for the probabilities Derive realization in classifier 11.

Pl = V1 (1-v0) p2 = 1-p1 w1 = p1f1 / (p1f1 + p2f2) q0 = v0.f2 / p1f1 + P2f2) q1 = w1 + v1.q0 In the following it is shown that f1 and f2 are measured and v0, v1 up to the initial values obtained by recursion from q0 and q1 can be.

Provided that vO, v11, f1 and f2 are known, you can all possible probabilities are given.

As explained in the previous sections, for the classification of measured values as object or terrain point the densities f1 (object-target) and f2 (Terrain) can be measured. This can be done in different ways: The distribution densities are location-dependent, the distribution densities are determined independently of location.

The distributions themselves, as a function of location or not, can be determined parametrically or by counting (histogram).

Three aspects play a role in determining densities.

- The reduction of the data, - the most exact reproduction possible (description), - the simple computational (hardware) representation.

f2, the brightness density of the terrain, is expediently used as location-dependent distribution function determined, since the terrain is generally stationary and the location dependency therefore contains very important information.

This is different with f1, the brightness density of the target.

Since the target usually moves and changes shape, it is Described by a location-independent distribution function, e.g. as a histogram.

For the acquisition of measured values 10, the brightness densities can only be calculated when a structure is adopted.

A Gaussian distribution can be assumed for the density f2 (terrain). The unknown parameters> i and & must be estimated or learned.

The equations for r, o are: r (k + 1) = s (X + 1) = x = current image r = reference image s = scatter image k = time, <1 learning factors An estimation equation can also be defined for f1.

This is necessary because f1 can change.

f. (k + 1) = f. (k) + γ. (f. (k) -f. (k)). u-1 u = 1 if the measured value with this coordinate in time step k it was classified as a terrain point u = 0 if the measured value with this coordinate in time step k is classified as a target point became.

γ <1 learning factor This means that all are necessary for the above equation Elements calculated.

The determination of the target coordinates in the target coordinates determiner is done as follows.

As already said, q is made up of different probabilities together and a probability, it is the presence probability, is used to calculate the target coordinates. For the calculation of the target coordinates In principle, there are several possibilities, the focus here was the target point the probability of presence chosen.

The coordinates of the target center of gravity can be used to describe the destination. This is obtained, based on the coordinates of the points classified as objects, by forming the moments of the 0th and 1st order. To this end, a target image Z (i, j) is first formed from the presence probability q1 with q5 as a selectable threshold value # 0 for q1 (i, j) # qs terrain point Z (i, j) = 1 for q1 (i, j)> q5 object point Line and column sums of object points can now be specified in the following way: i, j: Height, width of the target window The moments of the 0th and 1st order in the i and j directions are then: The center of gravity of the object results as follows: M1i Si M 0i M1j Sj = M0j With the filter 14 not only the target coordinates are smoothed, but also the future location of the target's presence is predicted with the help of a movement model, since this is not with Bayes -Theorem is possible, but is necessary for the calculation of the "a priori" probability from the "a posteriori" probability.

The calculation of new "a priori" probabilities from the previous ones. The calculated la-posteriorit probabilities are carried out in the delay element 13. Since adaptive (learning) processes work recursively, de "a-posteriori" - information must work back on the "a priori" information of the next time step. This results in the simple relationship: v0 (i, j, k + 1) = q0 (i, j, k) k: time i, j: location coordinates The equation is not without certain precautionary measures applicable because it assumes that the process is stationary.

For the foreground and thus for f2 and q0 there is no difficulty because the foreground is stationary and does not change its shape.

It is different with the target, which both moves and can change in its shape and size.

This means that both the position and the change in shape must be predicted will. There are several ways to solve this problem. If one assumes that the position of the target can be accurately estimated with a Kalman filter and that the change in shape is slow, then v1 (i, j, k + 1) = q1 (i- # i, j- # j, k) where #i and #j are the difference between the predicted target center of gravity S and measured center of gravity S is ~ ^ ~ 5. = i i #j = Sj-Sj the end The foregoing shows that the most essential advantage of using Bayes' theorem consists in the fact that the classification for each pixel is independent of all others Pixels happens. All measured values can be in sequential order as they are Sensor supplies can be classified immediately, without the measurement of a complete Image must be awaited.

Since one evaluates the complete information of the picture, and with it knows not only the probabilities for the target but also of the terrain, disturbances in the terrain are known and thus lead to far fewer misclassifications. The implementation does not cause any difficulties for a pixel clock of = 5 MHz.

L e r s e i t e

Claims (5)

Device for recognizing and tracking a target. PATENT CLAIMS 1. Device for the detection and tracking of a target, which one to electromagnetic Has radiation responsive sensor, thereby g e k e n n-characterized, a) that the Measured values supplied by the sensor are recorded in a measuring value acquisition (10) and on two Brightness densities for target (f1) and terrain (f2) reduced. b) that a classifier (ii) therefrom these as input variables (f1, f2) can be supplied, the classifier (ii) using Banes' theorem makes the classification in target and terrain and the measured value acquisition (10) over a switching function (u) controls, c) that the .l-posteriori signal (q) of the classifier (11) is fed to both a destination coordinate determiner (12) which contains the destination coordinates based on a focus on the probability of presence the target point is applied, determine 1 d) as well as one Delay element (13) from the a-posteriori signal (q) to the a-priori information retroactive signal (v) produces e) and that a filter (14) the target coordinates filters over time and the delay element (13) a difference signal (i, bj) between the predicted target center of gravity (s) and the measured center of gravity (s). 2. Device according to claim 1, characterized in that the measured values of the sensor are reduced so that the two brightness densities (f11 f2) are obtained as input variables, these being determined by: with the determining equations for r, #: r (k + 1) = s (k + 1) = s (k) + ß. (d (k) -s (k)). u where: x = current image r = reference image s = scatter image k = time. α, ß <1 learning factors A f1 (k + 1) = f1 (k) + γ. (f1 (k) -f1 (k)). | u-1 | f1 = estimated target histogram fi = measured target histogram u = 1, if the measured value with this coordinate in time step k was classified as a terrain point u = 0, if the measured value with this coordinate in time step k was classified as a target point. γ <1 learning factor 3. Device according to claim 1 and 2, characterized in that the classifier working according to Bayes' theorem (11) is designed so that it has a posteriori probability (wi) for the measurability of a process i according to the formula determined, where Pi and Wi are determined by: p1 = v1. (1-v0) P2 = 1-p1 w1 = p1f1 / (p1f1 + p2f2), w2 = 1-w1 q0 = v0.f2 / (p1f1 + p2f2) q1 = W1 + V1 q ° with p1.a-priori visibility probability for object p2.a-priori visibility probability for terrain (foreground and background) w1.a-posteriori visibility probability for object v0 a-priori presence probability for foreground v1 a-priori Probability of presence for object q0.a-posteriori probability of presence for foreground q1 a-posteriori probability of presence for object. 4. Device according to claims 1 to 3, characterized ge -kennz eichnet that the target coordinate determiner (12) is designed so that to determine the target coordinates using the coordinates of the target center of gravity, the moments 0 and 1st order are formed and initially a target image (Z (i, j)) is created from the probability of presence (q1) with a selectable threshold value (qs), with: 0 for q1 (i, j) #qs terrain point Z (i, j) = # 1 for q1 (i, j)> qs object point and from that row and column sums of object points according to: imax, jmax: Height, width of the target window, the moments 0 and 1 in the i and j directions are: The object's center of gravity results from: M1i Si = M0i M M1j Sj = M0j 5. Establishment according to claims 1 to 4, characterized in that the delay element (13) is designed to determine the new a priori probabilities become according to: v0 (i, j, k + 1) = q0 (i, j, k) k: time i, j: position coordinates and that using the values #i and #j supplied by the filter (14) is set: v1 (i, j, k + 1) = q1 (i- # i, j- # j, k) where #i and #j are the difference between the predicted Target center of gravity S and measured center of gravity S is #i = Si-Si #j = Sj-Sj