TW201520907A - Vision-based cyclist and pedestrian detection system and the methods - Google Patents

Abstract

This invention is a vision-based cyclist and pedestrian detection system. It is an apparatus for driver assistance system, including a video capture unit and an image processing unit. The video capture unit grab the image sequence and transfer the video to the image processing unit. The processing unit can extract the first order information including the histogram of oriented gradients (HOG) and the local oriented pattern (LOP). After that, it can extract the second order information including color self similarity and the texture self similarity information. By using the first order information and the second information, the system can verify if there is any cyclist or pedestrian in front of the host vehicle.

Description

Video camera treadmill and pedestrian detection system and method thereof

The present invention is an image-based treadmill and pedestrian detection system and method thereof, and more particularly, an image detection system for recognizing pedestrians and vehicles and a method thereof.

When driving, drivers should rely on direct visual inspection or through rearview mirrors to understand the road conditions of the next-door vehicles to ensure safe driving. Image and analysis are used for human and vehicle detection. Figure 1 is a conventional image detection system. As shown in the figure, the image capturing device 2 is disposed on the rear view mirror 21, and the image capturing device 2 is disposed on the rear view mirror 21, and the image capturing device can also be disposed on the driving and the deputy. On the window of the seat, take the view of the lane on the side of the vehicle. The detection methods in the prior art can be mainly divided into two categories: single frame-based and motion-based methods, in which a single image detection method is used to find a vehicle in an image. Features that differ from roads, defining areas of the image that may be vehicles, such as using the distribution of Entropy in the image, establishing a Saliency map, analyzing the distribution of edges in the image, or using The Circular template determines the area where the vehicle may be present.

The motion detection method mainly learns the image graphics caused by the motion of the vehicle. For example, the stereo camera (Stereo Camera) continuously detects the moving objects in the image, and establishes the similarity of the images obtained by the two cameras at the same time, and then establishes Similar to the image acquired in the next time, it is finally determined which object is an independently moving object. However, the single image detection method capacity It is susceptible to noise in the image and cannot accurately detect the presence of the vehicle and the vehicle in the image. The moving vehicle detection method, because the vehicle is in motion, causes the relative movement direction of the background and the vehicle to be indefinite, which makes the vehicle movement model difficult to manufacture.

In order to improve the recognition of image systems, conventional techniques can be divided into two main categories: one is to convert the problem of pedestrian detection into a model comparison problem, mainly by constructing human figures of different angles and postures, and then by ratio. The way to detect pedestrians, Gavrila et al. and Liu et al. use the human silhouette or edge image to represent the human form, and the human form is converted to DT. (Distance Transform) image. In order to more effectively overcome the changes of the object's translation, scale and orientation, Oren used Harr Vertical and Horizontal wavelets to calculate the humanoid feature map of the Wavelet Coefficients. In the study, the Histogram of Oriented Gradients is used to represent the characteristics of the human form. Through the SVM (Supported Vector Machine) machine learning method, the resulting classifier can effectively represent such a classifier. Features and detection as pedestrians in the image. Dalal and Triggs analyze and discuss the effects of different feature points applied to pedestrian detection, including HOGs (Histogram of Oriented Gradients), Haar Wavelets, PCA-SIFT and Shape Context. The results show that HOG can overcome pedestrian appearance. Mutation, and achieve good detection results. Walk et al. further proposed Color Self Similarity to complement the similarity parts not described by HOG. Wang et al. proposed combining HOG and LBP to further enhance the effect of pedestrian detection.

In view of the above disadvantages, the present invention provides an image-based treadmill and pedestrian detection system and a method thereof. The existing image-based driving safety devices for pedestrian detection have various types, and the design and manufacture are also different. However, most of them are not included in the detection of the motorcycle rider. In order to achieve effective detection of pedestrians and The treadmill rider, the knight and the pedestrian detection system and its method are important technical issues that need to be solved urgently for the current development of advanced vehicle technology.

In view of the above disadvantages of the prior art, the main object of the present invention is to provide an image-based treadmill and pedestrian detection system and a method thereof, which extract features from images, make appropriate groups of these features, and select and generate traits. The test results provide the driver's front lane condition to improve the safety of the driving.

Another object of the present invention is to provide an image treadmill and pedestrian detection system and method thereof, which utilize a direction gradient bar graph (HOG) and a local direction pattern feature (LOP feature) to support The Vector Machine (SVM) detects possible pedestrians, motorcyclists, and treadmills. Then, the information of the relative positional relationship between the motorcycle rider and the motorcycle is further utilized, and the filtering target may be detected to determine whether it is a pedestrian and a motorcycle rider, thereby achieving the purpose of the system.

The local direction pattern feature (LOP) generation method is similar to the local area bit pattern feature Local Binary Pattern. In addition to LOP, the calculation of the weight is added. The formula is as follows:

The LOP calculation can reinforce the weight of the material that changes significantly. As shown in Figure 3, if the flat image area that LBP may not be able to distinguish is simply separated, it can be separated by the weight of formula (1). The statistical area calculated by LOP is the same as LBP, which is an image area of 16 pixels by 16 pixels. 32 non-overlapping areas can be arranged in the image of 64 by 128, as shown in FIG. For each 16 by 16 total 256 possible positions, calculate the mode (with the LBP method) and the weight to which they belong to vote for the LOP feature. value.

In order to achieve the above object, the present invention provides an image treadmill and pedestrian detection system, which includes an image capture unit and an image processing unit, wherein the image capture unit is configured to generate a target image. And transmitting the target image to the image processing unit, and performing the first extraction and the second extraction on the target image through the image processing unit to generate a corresponding first-order image feature and a second-order image feature. The second-order image features include Color self similarity and Texture self similarity, which are obtained by calculating the similarity of the first-order image features of each image block. Through the machine learning method, through the classifier unit, it can be determined whether the first-order image signal and the second-order image feature are targets, and the purpose of the image recognition object is achieved. The target object consists of a cyclist, vehicle or pedestrian.

2‧‧‧Selection device

21‧‧‧ Rearview mirror

31‧‧‧Image capture unit

32‧‧‧Image Processing Unit

33‧‧‧ classifier unit

M1‧‧‧ target image

M2‧‧‧ first-order image features

M3‧‧‧ second-order image features

X1‧‧‧ first pick

X2‧‧‧Second pick

S1~S3‧‧‧Video Treadmill and Pedestrian Detection System Signal Flow Steps

Figure 1 is a schematic diagram of a conventional image object detection system.

2 is a schematic diagram of a method for detecting a treadmill and a pedestrian in the present invention.

Figure 3 is a schematic diagram of the image type treadmill and pedestrian detection system of the present invention.

Figure 4 is a flow chart of the signal type treadmill and pedestrian detection system signal of the present invention.

The embodiments of the present invention are described below by way of specific examples, and those skilled in the art can understand the other advantages and advantages of the present invention from the disclosure.

Please refer to FIG. 2 , which is a schematic diagram of a method for detecting a treadmill and a pedestrian in the present invention. As shown in the figure, step 1 (S1): capturing a target image. Example 2: Step 2 (S2): performing a first extraction according to the target image to generate a corresponding first-order image feature, wherein the first tiling system utilizes a direction gradient bar graph feature (HOG), region Directional feature (LOP), as the main way to identify the target image, obtain the image contour and texture of the target image; Step 3 (S3): Receive the first-order image feature, and the first-order image feature Performing a second extraction, which includes Color self similarity and Texture self similarity, and produces second-order image features by determining the order of the target image. Image features and second-order image signals achieve the purpose of image recognition.

Please refer to FIG. 3 , which is a schematic diagram of the camera treadmill and pedestrian detection system of the present invention. As shown in the figure, the system includes an image capturing unit 31 and an image processing unit 32. Please refer to FIG. The image capturing system is a flow chart of the image detecting system of the present invention. As shown in the figure, the image capturing unit 31 is configured to generate a target image M1, and transmit the target image M1 to the image processing unit 32, through the image processing unit. 32. Perform a first extraction of the target image M1 by X1. The first extraction of the X1 system utilizes a direction gradient bar graph feature (HOG) and a region direction feature (LOP) as the main way of identifying the target image. The image contour and texture of the target image generate a first-order image feature M2, and then the second-order image feature M2 is subjected to a second extraction of X2, and a corresponding second-order image signal M3 is generated. The first-order image feature M2 and the second-order image feature M3 are determined to include color self-similarity and texture self-similarity, and then passed through the classifier unit 33 to distinguish whether the image appears in the image. The object is to achieve the purpose of image recognition. The target may be a motorcycle rider or its vehicle or pedestrian.

The above-described embodiments are merely illustrative of the features and functions of the present invention, and are not intended to limit the scope of the technical content of the present invention. Any person skilled in the art can modify and change the above embodiments without departing from the spirit and scope of the present invention. Therefore, the scope of protection of this creation The scope shall be as listed in the scope of application for patents mentioned later.

31‧‧‧Image capture unit

32‧‧‧Image Processing Unit

M1‧‧‧ target image

M2‧‧‧ first-order image features

M3‧‧‧ second-order image features

X1‧‧‧ first pick

X2‧‧‧Second pick

Claims (10)

An image-based treadmill and pedestrian detection method, the method comprising: capturing a target image; performing a first extraction according to the target image to generate a first-order image feature; and receiving the first-order image feature, The second extraction of the first-order image feature is performed, and the second-order image feature is generated, and the image recognition object is achieved by determining the image feature of the target image and the second-order image feature and a classifier unit. The video camera treadmill and pedestrian detection method according to claim 1, wherein the first plucking system comprises a direction gradient bar graph feature (HOG) and a regional direction feature (LOP). The video camera treadmill and pedestrian detection method according to claim 1, wherein the first sculpt is used to distinguish an image contour and a texture in the target image. For example, the image-based treadmill and pedestrian detection method described in claim 1 wherein the second plucking system includes two features, namely, color self similarity and texture self-similarity. The sacredness of texture self similarity. The image-based treadmill and pedestrian detection method according to claim 4, wherein the second extraction further comprises the following steps: color self similarity calculation and texture self-similarity ( Texture self similarity) calculates an image treadmill and pedestrian detection system, which is included. An image capturing unit is configured to capture a target area and generate a target image; an image processing unit receives the target image, and performs a first extraction of the target image, and a second extraction, borrowing A first-order image feature and a second-order image feature of the target image are obtained. A classifier unit receives the image feature to distinguish whether a target object is present. The video camera treadmill and pedestrian detection system according to claim 6, wherein the first plucking system comprises a direction gradient bar graph feature (HOG) and a regional direction feature (LOP). The video camera treadmill and pedestrian detection system of claim 6, wherein the first extraction is used to distinguish image contours and textures within the target image. For example, the image type treadmill and pedestrian detection system described in claim 6 wherein the second extraction system has two feature extractions, including color self similarity extraction and texture. Texture self similarity is captured. The video camera treadmill and pedestrian detection system of claim 9, wherein the second extraction further comprises the following steps: color self-similarity calculation, texture self-similarity calculation.