Vehicle re-identification method based on group similarity

Abstract

The invention provides a vehicle re-identification method based on group similarity, and the method comprises the steps: selecting a proper group size n, and selecting a target vehicle and n-1 vehicles which are closest to the target vehicle under a first camera as a target group; calculating the similarity between the target group and all candidate groups under the second camera, and taking the group with the highest similarity as a matching group; and according to the group similarity of the matched groups, finding matched target individuals in the matched groups. According to the method, the vehicles with similar appearances can be distinguished by fully utilizing the information of the adjacent vehicles through the group similarity; according to the invention, by comparing the feature information of the vehicle group, the target vehicle can be identified more accurately; the method does not depend on the specific arrangement sequence of the individuals in the group, and is very suitable for the situation that the relative position change of the processing vehicle passing through the two adjacent cameras is small.

Description

Technical Field

[0001] The invention relates to the field of traffic technology, and in particular to a vehicle re-identification method based on group similarity. Background Art

[0002] With the acceleration of urbanization and the growth of traffic volume, the importance of vehicle re-identification technology in intelligent transportation systems has become increasingly prominent. By accurately identifying and tracking the same vehicle from different surveillance camera perspectives, it helps to improve traffic supervision efficiency and prevent and combat criminal behavior. In addition, this technology can connect the information captured by multiple cameras and realize the interconnection of information between multiple cameras, which is of great significance for realizing the intelligence of video surveillance systems.

[0003] In recent years, the rapid development of deep learning technology has provided new ideas for vehicle feature extraction. Deep neural network models have been proven to have powerful feature extraction capabilities and perform well on many computer vision tasks. However, vehicle re-identification technology based on vehicle appearance still faces many challenges, such as low image resolution, occlusion, perspective differences, and changes in lighting conditions. Low image resolution is a common problem, especially in long-distance vehicle images captured by surveillance cameras. Low-resolution images often cannot provide enough detailed information, making vehicle feature extraction difficult. In traffic scenes, vehicles may also be blocked by other vehicles, pedestrians, trees and other objects, resulting in some vehicle features not being captured. In this case, the accuracy of the vehicle re-identification system will be greatly affected. In addition, the data of vehicle appearance has the characteristics of high inter-class similarity and large intra-class differences. Different vehicles may have similar appearance features due to similar designs or colors, that is, high inter-class similarity, which makes it more difficult to accurately identify specific vehicles in massive vehicle data. However, due to the different camera installation positions and angles and changes in lighting conditions, images of the same vehicle at different times, locations, and angles may show different appearance features, that is, large intra-class differences. This further increases the difficulty of vehicle re-identification based on vehicle appearance. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a vehicle re-identification method based on group similarity. The present invention can make full use of the information of adjacent vehicles to distinguish vehicles with similar appearances through group similarity; and by comparing the characteristic information of the vehicle group, the target vehicle can be identified more accurately; and the present invention does not rely on the specific arrangement order of individuals in the group.

[0005] The technical solution of the present invention is: a vehicle re-identification method based on group similarity, comprising the following steps:

[0006] S1), construct a target group, select an appropriate group size n, and select the target vehicle and the n-1 vehicles closest to the target vehicle under the first camera as the target group;

[0007] S2), group matching, calculating the similarity between the target group and all candidate groups under the second camera, and taking the group with the highest similarity as the matching group;

[0008] S3), according to the group similarity of the matching group, find the matching target individual in the matching group.

[0009] Preferably, in step S1), the group size n is determined according to the change in front-to-back position of the vehicle between the first camera and the second camera.

[0010] Preferably, in step S1), the target vehicle is determined by the vehicle image captured by the first camera, and within a certain time range, the vehicle image captured by the second camera is obtained, and then arranged in chronological order, and all vehicle images captured within the time period are searched to determine the time when the target vehicle passes the second camera.

[0011] Preferably, in step S1), after determining the target vehicle, the vehicles are sorted in the order in which they pass the first camera, and the (n-1) / 2 vehicles in front of the target vehicle and the (n-1) / 2 vehicles behind the target vehicle are taken as the n-1 vehicles closest to the target vehicle.

[0012] Preferably, in step S2), for the target group A={a1, a2, a3, ..., a n} and candidate group B = {b1,b2,b3,…,b n}; where a i represents the i-th vehicle in target group A; b j represents the jth vehicle in the candidate group; P(A,B) represents the similarity between the target group A and the candidate group B.

[0013] Preferably, in step S2), the similarity P(A, B) between the target group A and the candidate group B is expressed as:

[0014]

[0015]

[0016]

[0017] x ij ∈(0,1)

[0018] In the formula, xij represents the matching relationship between the i-th individual in the target group A and the j-th individual in the candidate group B; p ij It represents the similarity between the i-th individual in the target group A and the j-th individual in the candidate group B.

[0019] Preferably, in step S2), if x ij =1, it means that the i-th individual in the target group A matches the j-th individual in the candidate group B; if x ij =0, it means that the i-th individual in population A does not match the j-th individual in population B.

[0020] Preferably, in step S2), the second camera and the first camera are adjacently arranged cameras, and the vehicle passes the first camera first and then the second camera.

[0021] Preferably, in step S3), according to the similarity of the matching groups, a matching target individual is found in the matching groups, that is, when x ij =1, indicating that the i-th individual in population A matches the j-th individual in population B.

[0022] The beneficial effects of the present invention are:

[0023] 1. The present invention can make full use of the information of adjacent vehicles to distinguish vehicles with similar appearances through group similarity; when the appearances of two vehicles are very similar, their group characteristics may be quite different because their adjacent vehicles may be different, and the present invention can more accurately identify the target vehicle by comparing the characteristic information of the vehicle group;

[0024] 2. The method of the present invention does not depend on the specific arrangement order of individuals in the group. The present invention is very suitable for processing the situation where the relative position of a vehicle passing two adjacent cameras changes slightly. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic diagram of the process of the present invention;

[0026] Figure 2 is a visualization result of the similarity matrix in an embodiment of the present invention;

[0027] Figure 3 This is a curve diagram showing the impact of the relative position change of vehicles on the group similarity method in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings:

[0029] like Figure 1 As shown, this embodiment provides a vehicle re-identification method based on group similarity, comprising the following steps:

[0030] S1), construct a target group, select an appropriate group size n, and select the target vehicle and the n-1 vehicles closest to the target vehicle under the first camera as the target group;

[0031] S2), group matching, calculating the similarity between the target group and all candidate groups under the second camera, and taking the group with the highest similarity as the matching group;

[0032] S3), according to the group similarity of the matching group, find the matching target individual in the matching group.

[0033] As a preferred embodiment of the present invention, in step S1), the group size n is determined according to the front-rear position change of the vehicle passing the first camera and the second camera.

[0034] As a preferred embodiment of the present invention, in step S1), the target vehicle is determined by the vehicle image captured by the first camera, and within a certain time range, the vehicle image captured by the second camera is obtained, and then arranged in chronological order, and all vehicle images captured within the time period are searched to determine the time when the target vehicle passes the second camera.

[0035] As a preferred embodiment of the present invention, in step S1), after determining the target vehicle, the vehicles are sorted in the order in which they pass the first camera, and the (n-1) / 2 vehicles in front of the target vehicle and the (n-1) / 2 vehicles behind the target vehicle are taken as the n-1 vehicles closest to the target vehicle.

[0036] As a preferred embodiment of this embodiment, in step S2), based on the time when the target vehicle appears in the first camera, the speed of the target vehicle, and the distance between the first camera and the second camera, the time period in which the target vehicle passes the second camera is estimated, and the video of this time period is captured, all vehicles in this time period are selected as preferred vehicles, and the candidate group is divided.

[0037] As a preferred embodiment of this invention, in step S2), for the target group A={a1, a2, a3, ..., a n} and candidate group B = {b1,b2,b3,…,b n}; where a i represents the i-th vehicle in target group A; b j represents the jth vehicle in the candidate group; P(A,B) represents the similarity between the target group A and the candidate group B; the similarity P(A,B) between the target group A and the candidate group B is expressed as:

[0038]

[0039]

[0040]

[0041] x ij ∈(0,1)

[0042] In the formula, x ij represents the matching relationship between the i-th individual in the target group A and the j-th individual in the candidate group B; p ij It represents the similarity between the i-th individual in the target group A and the j-th individual in the candidate group B.

[0043] As a preferred embodiment of this invention, in step S2), if x ij =1, it means that the i-th individual in the target group A matches the j-th individual in the candidate group B; if x ij =0, it means that the i-th individual in population A does not match the j-th individual in population B.

[0044] As a preferred embodiment of the present invention, in step S2), the second camera and the first camera are adjacently arranged cameras, and the vehicle first passes through the first camera and then passes through the second camera.

[0045] As a preferred embodiment of this embodiment, in step S3), a matching target individual is found in the matching group according to the similarity in the matching group, that is, when x ij =1, indicating that the i-th individual in population A matches the j-th individual in population B.

[0046] In order to verify the accuracy of the method of this embodiment, this embodiment is compared with the traditional method. This embodiment sets the list length N=50, 100, 200 respectively, where the list length N represents the range of vehicle search. At the same time, this embodiment also sets different group sizes.

[0047] When the vehicle passes through two cameras and the relative position between the vehicles remains unchanged, the accuracy of the method of this embodiment is much higher than that of the traditional method based on individual similarity re-identification method. When the size of the group is selected appropriately, whether the number of vehicles is 50, 100 or 200, the accuracy of the method of this embodiment is 0.92, 0.9 and 0.91 respectively; while the accuracy of the traditional method based on individual similarity re-identification method is 0.44, 0.31 and 0.22 respectively; the method of this embodiment is significantly higher than the traditional method. In addition, the method of this embodiment shows better robustness to changes in list length. When the list length increases, the accuracy of the traditional method continues to decrease, while the accuracy of the method based on this embodiment decreases relatively slowly overall, showing a stable high accuracy. Because when the list length increases, the probability of two or more similar groups appearing in the list is much smaller than the probability of two or more similar individuals appearing in the list.

[0048] In order to intuitively demonstrate the principle of the group similarity method, this embodiment visualizes the similarity matrix P between the two lists, where P(i,j) represents the similarity between the i-th group (or individual) in list 1 and the j-th group (or individual) in list 2. List 1 represents the vehicle images captured by the first camera in chronological order, and list 2 represents the vehicle images captured by the second camera in chronological order. When the relative position remains unchanged, the visualization result of the similarity matrix P between the two lists is as follows: Figure 2 As shown, where n represents the size of the group.

[0049] Figure 2 The depth of the midpoint color represents the size of P(i,j). The darker the color, the higher the similarity between the two groups (or individuals). The horizontal and vertical coordinates represent the position of the group (or individual) in the two lists. Because the relative position between the vehicles does not change at all when the vehicle passes through the two cameras, the value on the diagonal represents the similarity between the same group (or individual) in the two lists. When the individual similarity (n=1) of the traditional method is used, the accuracy of re-identification is not high due to the large number of similar individuals in the list. When the group similarity (n>1) of the method of this embodiment is used, since the probability of two or more similar groups appearing in the list is much smaller than the probability of two or more similar individuals appearing in the list, the use of group similarity can greatly improve the accuracy of re-identification. The larger the group, the smaller the probability of two or more similar groups appearing in the list, and the larger the value on the diagonal. However, as the group increases, the probability of two or more similar individuals appearing in the group also increases, resulting in a decrease in the accuracy of re-identification. In addition, the increase in the group also means an increase in the complexity of group similarity calculation. Therefore, the group cannot be too large.

[0050] At the same time, this embodiment also changes the relative positions of the vehicles when they pass through the two cameras. This embodiment keeps the positions of the vehicles in list 1 unchanged and adjusts the positions of the vehicles in list 2. The specific adjustment method is as follows: Assume s i ~N(i,σ 2 ),i∈{0,1,…,N-1}, for s i Perform a set of sampling and get [s′0,s′1,…,s′ N-1 ]Let f be a value that satisfies s′ f(0) <s′ f(1) <… <s′ f(N-1) reversible mapping, and f(i)∈{0,1,…,N-1}, let y i =f -1 (i) This real-time basis i Adjust the position of the vehicles in List 2. i =j means the label in list 1 is x i The adjusted position of the vehicle in list 2 is j.

[0051] By using different standard deviations σ, the positions of the vehicles in List 2 can be adjusted to different degrees. The influence of the relative position change (which can be quantified by σ) when the vehicle passes the two cameras on the experimental results is shown in the figure. Figure 3 As shown, the curve with n=1 represents a method based on individual similarity (traditional method), and the curve with n>1 represents a method based on group similarity (the method of this embodiment).

[0052] from Figure 3 It can be seen that using different standard deviations σ has no effect on the vehicle re-identification method based on individual similarity (group size is 1). This is because the calculation of individual similarity is independent of the relative position change between vehicles when they pass through two cameras. As the standard deviation σ increases, the change in the relative position between vehicles when they pass through two cameras also gradually increases. The effect of the method based on group similarity gradually decreases, and the rate of decrease is related to the size of the group. In general, the larger the group, the slower the rate of decrease, and the better the robustness to the relative position change between vehicles when they pass through two cameras. This is determined by the calculation method of group similarity. Because the similarity between two groups is independent of the arrangement order of individuals within the group, the larger the group, the smaller the impact of the relative position change between vehicles on the group similarity method. When σ<2, the accuracy of the method based on group similarity is generally higher than that of the method based on individual similarity. When σ is large, the group size needs to be appropriately increased.

[0053] The above embodiments and descriptions are only for illustrating the principles and best embodiments of the present invention. Without departing from the spirit and scope of the present invention, the present invention may be subject to various changes and improvements, all of which fall within the scope of the present invention to be protected.

Claims

1. A vehicle re-identification method based on group similarity, characterized in that: The following steps are involved: S1), construct a target group, select an appropriate group size n, and select the target vehicle and the n-1 vehicles closest to the target vehicle under the first camera as the target group; S2), group matching, calculating the similarity between the target group and all candidate groups under the second camera, and taking the group with the highest similarity as the matching group; S3), according to the group similarity of the matching group, find the matching target individual in the matching group.

2. The vehicle re-identification method based on group similarity according to claim 1, characterized in that: In step S1), the group size n is determined according to the change in the front-rear position of the vehicle between the first camera and the second camera.

3. The vehicle re-identification method based on group similarity according to claim 2, characterized in that: In step S1), the target vehicle is determined by the vehicle image captured by the first camera, and the vehicle image captured by the second camera is obtained within a certain time range, and then arranged in chronological order, and all vehicle images captured within the time period are searched to determine the time when the target vehicle passes the second camera.

4. The vehicle re-identification method based on group similarity according to claim 3, characterized in that: In step S1), after the target vehicle is determined, the order in which the vehicles pass the first camera is sorted, and the (n-1) / 2 vehicles in front of the target vehicle and the (n-1) / 2 vehicles behind the target vehicle are regarded as the n-1 vehicles closest to the target vehicle.

5. The vehicle re-identification method based on group similarity according to claim 1, characterized in that: In step S2), based on the time when the target vehicle appears in the first camera, the speed of the target vehicle, and the distance between the first camera and the second camera, the time period in which the target vehicle passes the second camera is estimated, and the video of this time period is captured. All vehicles in this time period are selected as preferred vehicles and divided into candidate groups.

6. The vehicle re-identification method based on group similarity according to claim 5, characterized in that: In step S2), for the target group A = {a1, a2, a3, ..., a n } and candidate group B = <h2 style=";text-align:left;direction:ltr">{b1,b2,b3,…,b<h2 style=";text-align:left;direction:ltr"> n <h2 style=";text-align:left;direction:ltr">}; Among them, a i represents the i-th vehicle in target group A; b j represents the jth vehicle in the candidate group; P(A,B) represents the similarity between the target group A and the candidate group B.

7. The vehicle re-identification method based on group similarity according to claim 6, characterized in that: In step S2), the similarity P(A, B) between the target group A and the candidate group B is expressed as: x ij ∈(0,1) In the formula, x ij represents the matching relationship between the i-th individual in the target group A and the j-th individual in the candidate group B; p ij It represents the similarity between the i-th individual in the target group A and the j-th individual in the candidate group B.

8. The vehicle re-identification method based on group similarity according to claim 7, characterized in that: In step S2), if x ij =1, it means that the i-th individual in the target group A matches the j-th individual in the candidate group B; if x ij =0, it means that the i-th individual in population A does not match the j-th individual in population B.

9. The vehicle re-identification method based on group similarity according to claim 5, characterized in that: In step S2), the second camera and the first camera are adjacently arranged cameras, and the vehicle passes the first camera first and then the second camera.

10. The vehicle re-identification method based on group similarity according to claim 8, characterized in that: In step S3), according to the similarity of the matching groups, the matching target individual is found in the matching groups, that is, when x ij =1, indicating that the i-th individual in population A matches the j-th individual in population B.

Images