Re-identification system

The re-identification system improves accuracy by incorporating orientation-based ranking rules to handle varying object orientations, enhancing precision in identifying the same object across multiple viewpoints.

JP2024166888AActive Publication Date: 2024-11-29TOYOTA JIDOSHA KK
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
JP2023083290
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-19
Publication Date
2024-11-29
Estimated Expiration
2043-05-19

AI Technical Summary

Technical Problem

Existing re-identification systems face accuracy issues due to varying orientations of moving objects captured from different viewpoints, which complicates the re-identification process.

Method used

A re-identification system that considers the orientation of moving objects by assigning ranks to similarity degrees based on matching orientations, using a ranking rule to improve accuracy through a multi-step determination process.

Benefits of technology

Enhances the accuracy of re-identification by accounting for object orientations, ensuring higher precision in identifying the same object across multiple videos.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024166888000001_ABST
    Figure 2024166888000001_ABST
Patent Text Reader

Abstract

To improve accuracy of re-identification of a moving object.SOLUTION: Feature information indicates a combination of a re-identification feature amount and a direction of a moving object in a video. A rank regarding a similarity between two moving objects is highest when the two moving objects are identical and the two moving objects have the same direction, and is lowest when the two moving objects are not identical and the two moving objects have different directions. A ranking rule is that the rank is higher as the similarity is higher. A re-identification system calculates a plurality of similarities between two moving objects based on the re-identification feature amount, and temporarily determines whether the two moving objects are identical. Further, the re-identification system determines a plurality of ranks for each of the plurality of similarities, and calculates a degree of consistency which is a degree that a correspondence between the plurality of similarities and the plurality of ranks is consistent with the ranking rule. Then, the re-identification system determines whether the two moving objects are identical based on the degree of consistency and the plurality of similarities.SELECTED DRAWING: Figure 3
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present disclosure relates to a technology for re-identification of a moving object. [Background technology]

[0002] Person re-identification is a technology that identifies the same person from multiple videos. A person is detected from the images that make up the video, and the overall features of the detected person are extracted, and person re-identification is performed based on the extracted features. The feature extraction is performed based on a person re-identification model based on machine learning.

[0003] As techniques relating to extraction of a person's feature amount and identification of the same person, for example, those disclosed in Patent Documents 1 to 4 are known. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2011-100175 A [Patent Document 2] International Publication No. 2014 / 132841 [Patent Document 3] JP 2017-097549 A [Patent Document 4] JP 2017-033547 A Summary of the Invention [Problem to be solved by the invention]

[0005] Consider the re-identification of moving objects (e.g. people, robots, vehicles). The apparent "orientation" of the same moving object captured in multiple images may differ. For example, the orientation of the same moving object may differ between multiple images captured by multiple cameras with different viewpoints. If the orientation of the same moving object differs significantly between multiple images, the accuracy of the re-identification process may decrease. A great deal of effort would be required to solve such problems by training a re-identification model alone.

[0006] An object of the present disclosure is to provide a technology capable of improving the accuracy of re-identification of a moving object. [Means for solving the problem]

[0007] One aspect of the present disclosure relates to a re-identification system for re-identifying a mobile object. The re-identification system is one or more processors; one or more storage devices for storing feature information indicating a combination of re-identification features and orientations of moving objects in a video; Equipped with. The rank of the similarity between two moving objects is The highest rank is when the two moving objects are identical and the orientations of the two moving objects are the same; The lowest rank is when the two moving objects are different and the orientations of the two moving objects are different. An intermediate rank between the highest and lowest ranks Includes. The ranking rule is that the higher the rank, the higher the similarity. The one or more processors Calculating a plurality of similarities between the first moving object and the second moving object based on a plurality of combinations of the re-identified feature amount of the first moving object and the re-identified feature amount of the second moving object; performing a first determination process for provisionally determining whether the first moving body and the second moving body are the same or not based on each of the plurality of similarities; determining a plurality of ranks for each of the plurality of similarities based on a result of the first determination process, an orientation of the first moving body, and an orientation of the second moving body; calculating a degree of correspondence between the plurality of similarities and the plurality of ranks and a ranking rule by comparing the correspondence between the plurality of similarities and the plurality of ranks and a ranking rule; A second determination process is performed to determine whether the first moving object and the second moving object are the same or not based on the degree of match and the multiple similarities. It is configured as follows. Effect of the Invention

[0008] According to the present disclosure, the re-identification process is performed taking into account the orientation of the moving body. More specifically, a rank is assigned to the similarity in the re-identification process taking into account the orientation of the moving body. Furthermore, a matching degree, which is the degree to which the ranking matches the ranking rule, is calculated. Then, the re-identification process is performed taking into account the matching degree. In this way, the re-identification process is performed taking into account the orientation of the moving body, thereby improving the accuracy of the re-identification process. [Brief description of the drawings]

[0009] [Figure 1] 1 is a conceptual diagram for explaining an overview of a ReID system according to an embodiment. [Diagram 2] 4 is a block diagram showing an example of a functional configuration related to a feature information acquisition process according to the embodiment; FIG. [Diagram 3] 4 is a block diagram showing an example of a functional configuration related to a re-identification process according to an embodiment. FIG. [Figure 4] 11A and 11B are diagrams for explaining an example of a re-identification process according to an embodiment; [Diagram 5] 11A and 11B are diagrams for explaining an example of a re-identification process according to an embodiment; [Figure 6] 11A and 11B are diagrams for explaining an example of a re-identification process according to an embodiment; DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] Embodiments of the present disclosure will be described with reference to the accompanying drawings.

[0011] 1. Overview of the re-identification system Moving object re-identification is a technology for identifying the same moving object from multiple videos. Examples of moving objects include people (pedestrians), robots, vehicles, etc. In the following explanation, as an example, we consider the case where the moving object is a person (Person Re-identification). When generalizing, "person" in the following explanation should be read as "moving object". Also, in the following explanation, re-identification may be called "ReID".

[0012] FIG. 1 is a conceptual diagram for explaining an overview of a ReID system 100 according to the present embodiment. The ReID system 100 re-identifies a person 1. More specifically, the ReID system 100 acquires video VIDs captured by one or more cameras 10 installed in a predetermined area. The ReID system 100 identifies the same person from multiple video VIDs. For example, the ReID system 100 identifies the same person 1 from multiple video VIDs captured by different cameras 10. As another example, the ReID system 100 identifies the same person 1 from multiple video VIDs captured at different times by the same camera 10. By such a re-identification process, it becomes possible to track the same person 1, for example.

[0013] The ReID system 100 may be configured with a single node or multiple nodes. More specifically, the ReID system 100 includes one or multiple processors 110 (hereinafter simply referred to as "processor 110"), one or multiple storage devices 120 (hereinafter simply referred to as "storage devices 120"), and an interface 130. The processor 110 executes various processes. For example, the processor 110 includes a CPU (Central Processing Unit). The storage device 120 stores various information required for the processes. Examples of the storage device 120 include a volatile memory, a non-volatile memory, a HDD (Hard Disk Drive), and an SSD (Solid State Drive). The interface 130 receives various data from the outside and outputs various data to the outside.

[0014] The processor 110 acquires a plurality of video VIDs via the interface 130. The plurality of video VIDs are captured by one or a plurality of cameras 10 installed in a predetermined area. The processor 110 performs a re-identification process based on the plurality of video VIDs. Then, the processor 110 outputs a result of the re-identification process via the interface 130.

[0015] The program PROG is a computer program for performing re-identification processing. The program PROG is stored in the storage device 120. The program PROG may be recorded in a computer-readable recording medium. The processing by the ReID system 100 (processor 110) may be realized by cooperation between the processor 110 that executes the program PROG and the storage device 120.

[0016] The "orientation" of the same person 1 may differ in appearance in multiple video VIDs. For example, the orientation of the same person 1 may differ between multiple video VIDs captured by multiple cameras 10 with different viewpoints. If the orientation of the same person 1 differs significantly between multiple video VIDs, the accuracy of the re-identification process may decrease.

[0017] Therefore, the ReID system 100 according to the present embodiment is configured to perform re-identification processing with high accuracy, taking into account the "orientation" of the person 1. The processing by the ReID system 100 (processor 110) according to the present embodiment will be described in more detail below.

[0018] 2. Feature information acquisition process 2 is a block diagram showing an example of a functional configuration related to feature information acquisition processing by the ReID system 100 (processor 110) according to this embodiment. The ReID system 100 includes, as functional blocks, a human detector 200, a tracker 210, a ReID feature extractor 220, a pose estimator 230, and a direction estimator 240.

[0019] A video VID is input to the person detection unit 200. The video VID includes a series of images (frames). The person detection unit 200 performs a person detection process to detect a person 1 in each image. A bounding box represents the position of the person 1 detected in the image. The person detection unit 200 acquires information on the bounding box of the person 1 in each image. Note that the person detection process is a well-known technique, and the method is not particularly limited. For example, YOLOX is used as the person detection unit 200.

[0020] The tracker 210 automatically tracks the same person 1i in a series of images based on a tracking algorithm. More specifically, the tracker 210 tracks the same person 1i in a series of images by tracking bounding boxes representing the same person 1i in the series of images. The tracker 210 associates multiple bounding boxes representing the same person 1i in multiple time steps (t1, t2, t3, . . . ). A "track TRi" is a set of multiple bounding boxes representing the same person 1i in multiple time steps (t1, t2, t3, . . . ). In other words, the track TRi is information representing the time sequence of the same person 1i in a series of images. Note that the subscript i is identification information of the track TRi, and is not identification information indicating who the person 1i is. In addition, the tracking process is a well-known technology, and the method is not particularly limited. For example, ByteTrack is used as the tracker 210.

[0021] The ReID feature extraction unit 220 extracts the feature of the person 1i for re-identification processing based on the image of the part of the person 1i. The feature of the person 1i for re-identification processing is hereinafter referred to as the ReID feature FEAi[t]. More specifically, the ReID feature extraction unit 220 acquires a partial image surrounded by each of the bounding boxes of the time steps (t1, t2, t3, . . .) constituting the track TRi. The partial image corresponds to the image of the person 1i. Then, the ReID feature extraction unit 220 extracts the ReID feature FEAi[t] from each partial image by using a ReID model based on machine learning. In this way, the ReID feature extraction unit 220 extracts the ReID feature FEAi[t] of the person 1i in each of the time steps (t1, t2, t3, . . .) based on the track TRi. The ReID feature extraction process is a well-known technique, and the method is not particularly limited. The ReID model may be a Transformer-based model.

[0022] The posture estimation unit 230 estimates the posture of the person 1i based on an image of a portion of the person 1i. More specifically, the posture estimation unit 230 acquires a partial image surrounded by each of a plurality of bounding boxes of a plurality of time steps (t1, t2, t3, . . .) constituting the track TRi. The partial image corresponds to the image of the person 1i. The posture estimation unit 230 extracts key points from each partial image by using a posture estimation model based on machine learning, and estimates the posture of the person 1i. In this way, the posture estimation unit 230 estimates the posture of the person 1i at each of a plurality of time steps (t1, t2, t3, . . .) based on the track TRi. Note that the posture estimation process is a well-known technique, and the method is not particularly limited. For example, TransPose is used as the posture estimation unit 230.

[0023] The orientation estimation unit 240 estimates the apparent orientation DIR[t] of the person 1i in the image based on the posture of the person 1i. More specifically, the orientation estimation unit 240 determines which of a plurality of orientations DIR1 to DIRn the orientation DIR[t] of the person 1i at each time step (t1, t2, t3, . . . ) is, based on the posture of the person 1i. Here, n is an integer of 2 or more. For example, when n=4, the orientation DIR[t] of the person 1i in the image is classified into DIR1 (front; -45 degrees to +45 degrees), DIR2 (rightward; +45 degrees to +135 degrees), DIR3 (leftward; -45 degrees to -135 degrees), and (backward; +135 degrees to +180 degrees, -135 degrees to -180 degrees). Note that n is not limited to 4. In this manner, the orientation estimation unit 240 estimates the orientation DIR[t] of the person 1i in each of the multiple time steps (t1, t2, t3, . . . ).

[0024] The feature information 250 is information indicating a combination of the ReID feature amount FEAi[t] and the direction DIR[t] of the person 1i in the video VID. In the above example, the feature information 250 indicates a combination of the ReID feature amount FEAi[t] and the direction DIR[t] for each track TRi. In other words, the feature information 250 indicates a correspondence relationship between the track TRi, the ReID feature amount FEAi[t], and the direction DIR[t]. Such feature information 250 is stored in the storage device 120. Various pieces of feature information 250 related to various tracks TRi of various video VIDs may be generated and updated as a ReID database. Such a ReID database is also stored in the storage device 120.

[0025] According to this embodiment, the feature information 250 indicating the combination of the ReID feature amount FEAi[t] and the direction DIR[t] is used in the re-identification process. The re-identification process using the feature information 250 will be described in detail below.

[0026] 3. Re-identification process Hereinafter, a re-identification process will be considered, which determines whether a first person 1-A appearing in a first video VID-A and a second person 1-B appearing in a second video VID-B are the same person. A first track TR-A is a track TR relating to a first person 1-A appearing in a first video VID-A. A second track TR-B is a track TR relating to a second person 1-B appearing in a second video VID-B. A first feature information 250-A is feature information 250 relating to a first person 1-A obtained based on a first track TR-A. A second feature information 250-B is feature information 250 relating to a second person 1-B obtained based on a second track TR-B.

[0027] 3 is a block diagram showing an example of a functional configuration related to the re-identification process by the ReID system 100 (processor 110) according to this embodiment. The ReID system 100 includes, as functional blocks, a similarity calculation unit 300, a score calculation unit 310, a determination unit 320, a ranking unit 330, and a verification unit 340.

[0028] The similarity calculation unit 300 acquires the first feature information 250-A related to the first person 1-A and the second feature information 250-B related to the second person 1-B from the storage device 120. The similarity calculation unit 300 performs a "similarity calculation process" to calculate the similarity SIM between the first person 1-A and the second person 1-B based on the first feature information 250-A and the second feature information 250-B. More specifically, the similarity calculation unit 300 calculates the similarity SIM between the first person 1-A and the second person 1-B based on a combination of the ReID feature amount FEA-A[t] of the first person 1-A and the ReID feature amount FEA-B[t] of the second person 1-B. The closer the distance between the ReID feature amount FEA-A[t] and the ReID feature amount FEA-B[t] projected in the feature amount space, the higher the similarity SIM.

[0029] FIG. 4 shows an example of the similarity calculation process. The first feature information 250-A includes x combinations of the ReID feature amount FEA-A[t] and the direction DIR[t] of the first person 1-A regarding the first track TR-A. The second feature information 250-B includes y combinations of the ReID feature amount FEA-B[t] and the direction DIR[t] of the second person 1-B regarding the second track TR-B. x and y are integers equal to or greater than 1, and at least one of x and y is multiple. The similarity calculation unit 300 calculates multiple similarities SIMj (j=1 to x×y) based on multiple combinations of the x ReID feature amount FEA-A[t] and the y ReID feature amount FEA-B[t].

[0030] Next, the score calculation unit 310 performs a "first score calculation process (provisional score calculation process)" to calculate a score SCRj for each similarity SIMj. More specifically, the score calculation unit 310 calculates a plurality of scores SCRj (j=1 to x×y) based on each of the plurality of similarities SIMj (j=1 to x×y). The higher the similarity SIMj, the higher the score SCRj corresponding to that similarity SIMj.

[0031] An example of the first score calculation process is shown in Fig. 4. In the example shown in Fig. 4, the weighted score SCRj is calculated taking into consideration whether the orientation DIR of the first person 1-A and the orientation DIR of the second person 1-B match. More specifically, the first weighting coefficient λ1 is a weighting coefficient when the orientation DIR of the first person 1-A and the orientation DIR of the second person 1-B match. On the other hand, the second weighting coefficient λ2 is a weighting coefficient when the orientation DIR of the first person 1-A and the orientation DIR of the second person 1-B do not match. The second weighting coefficient λ2 is smaller than the first weighting coefficient λ1 (λ1>λ2). The similarity SIMj is calculated based on a combination of the ReID feature FEA-A[t] and the ReID feature FEA-B[t], where the orientation DIR[t] of the first person 1-A in the time step of the ReID feature FEA-A[t] is obtained from the first feature information 250-A, and the orientation DIR[t] of the second person 1-B in the time step of the ReID feature FEA-B[t] is obtained from the second feature information 250-B. The score calculation unit 310 selects the first weighting factor λ1 or the second weighting factor λ2 as a weighting factor for each of the multiple similarities SIMj (j=1 to x×y) based on the orientation DIR[t] of the first person 1-A and the orientation DIR[t] of the second person 1-B. Then, as shown in FIG. 4, the score calculation unit 310 calculates the score SCRj by multiplying the similarity SIMj by the selected weighting factor.

[0032] As a modified example, the score calculation unit 310 may directly set the similarity SIMj as the score SCRj without weighting. In either case, the higher the similarity SIMj, the higher the score SCRj corresponding to that similarity SIMj.

[0033] Next, the determination unit 320 performs a "first determination process (provisional determination process)" to provisionally determine whether or not the first person 1-A and the second person 1-B are the same based on each of the multiple scores SCRj (j = 1 to x x y). More specifically, the determination unit 320 compares each score SCRj with a threshold value TH. If the score SCRj is equal to or greater than the threshold value TH, the determination unit 320 provisionally determines that the first person 1-A and the second person 1-B are the same. On the other hand, if the score SCRj is less than the threshold value TH, the determination unit 320 provisionally determines that the first person 1-A and the second person 1-B are not the same. Since the first determination process is performed for each score SCRj, x x y determination results are obtained.

[0034] Next, the ranking unit 330 performs a "ranking process" to rank each of the multiple similarities SIMj (j=1 to x×y).

[0035] An example of the definition of ranks and ranking process will be described with reference to Fig. 5. There are four ranks related to the similarity between two people (1-A, 1-B), Rank1 to Rank4. The definitions of Rank1 to Rank4 are as follows. Rank 1: The two people are the same and their orientations DIR are the same. Rank 2: The two people are the same, and the orientations DIR of the two people are different. Rank 3: The two people are different and the direction DIR of the two people is the same. Rank 4: The two people are different and the directions DIR of the two people are different.

[0036] The ranking order is Rank1>Rank2>Rank3>Rank4. That is, Rank1 is the highest rank and Rank4 is the lowest rank. Rank2 and Rank3 are intermediate ranks between Rank1 and Rank4.

[0037] The ranking unit 330 determines a plurality of ranks for each of the plurality of similarities SIMj (j=1 to x×y). The similarity SIMj is calculated based on a combination of the ReID feature amount FEA-A[t] and the ReID feature amount FEA-B[t], and the orientation DIR[t] of the first person 1-A in the time step of the ReID feature amount FEA-A[t] is obtained from the first feature information 250-A, and the orientation DIR[t] of the second person 1-B in the time step of the ReID feature amount FEA-B[t] is obtained from the second feature information 250-B. Whether the first person 1-A and the second person 1-B are the same in terms of the similarity SIMj is obtained from the result of the above-mentioned first determination process. Therefore, the ranking unit 330 can determine a plurality of ranks for each of the plurality of similarities SIMj (j=1 to x×y) based on the result of the first determination process, the first feature information 250-A, and the second feature information 250-B. As a result, as shown in FIG. 5, a correspondence relationship between a plurality of similarities SIMj and a plurality of ranks is obtained.

[0038] If it is normal, the similarity SIMj assigned with Rank1 should be higher than the similarity SIMj assigned with Rank2, which should be higher than the similarity SIMj assigned with Rank3, which should be higher than the similarity SIMj assigned with Rank4. In other words, if it is normal, the higher the rank, the higher the similarity SIMj should be.

[0039] From the above viewpoints, the verification unit 340 performs a "verification process" to verify whether or not the correspondence between the multiple similarities SIMj (j=1 to x×y) and the multiple ranks is valid. More specifically, the verification unit 340 holds, as a ranking rule 350, that "the higher the rank, the higher the similarity SIMj." The verification unit 340 compares the correspondence between the multiple similarities SIMj and the multiple ranks with the ranking rule 350 to verify whether or not the correspondence is valid.

[0040] 5 shows both normal examples and examples including an abnormality. In this example, the order of similarities is assumed to be SIM1, SIM2, . . . , SIM25. In the normal example, the correspondence between the similarities SIM1 to SIM25 and the ranks matches the ranking rule 350. On the other hand, in the example including an abnormality, the combination of similarity SIM3 and Rank3, and the combination of similarity SIM24 and Rank1 do not match the ranking rule 350.

[0041] The degree of agreement αj is the degree to which the correspondence between the multiple similarities SIMj (j=1 to x×y) and the multiple ranks matches the ranking rule 350. The verification unit 340 calculates the degree of agreement αj for each of the multiple similarities SIMj based on the result of the verification process. For example, the degree of agreement αj is set to take a value of 0 to 1. If the combination of the similarity SIMj and the rank matches the ranking rule 350, the degree of agreement αj is 1.0. On the other hand, if the combination of the similarity SIMj and the rank does not match the ranking rule 350, the degree of agreement αj is less than 1. The more the combination of the similarity SIMj and the rank deviates from the ranking rule 350, the lower the degree of agreement αj may be set. For example, if Rank2 is mistakenly assigned to a similarity SIMj that should be assigned normally, the degree of deviation is "1". If Rank4 is mistakenly assigned to a similarity SIMj that should be assigned normally, the degree of deviation is "3". As the discrepancy increases, the agreement αj approaches 0.

[0042] As a modified example, a common matching degree αj may be calculated for all of the multiple similarities SIMj (j=1 to x×y). For example, if there are two anomalies for 25 similarities SIM1 to SIM25, the common matching degree αj is 23 / 25.

[0043] It can be said that the degree of match αj obtained by the verification process represents the validity and reliability of the score SCRj calculated by the first score calculation process (provisional score calculation process). Therefore, according to this embodiment, the score SCRj is corrected taking into account the degree of match αj. For this purpose, the degree of match αj obtained by the verification process is fed back to the score calculation unit 310.

[0044] Upon receiving the degree of match αj, the score calculation unit 310 performs a "second score calculation process." In the second score calculation process, the score calculation unit 310 calculates a corrected score CSCRj by correcting the score SCRj based on the degree of match αj. An example of the second score calculation process is shown in Fig. 6. The score calculation unit 310 multiplies the original score SCRj for the similarity SIMj by the degree of match αj for the similarity SIMj to correct the original score SCRj and calculates a corrected score CSCRj.

[0045] The determination unit 320 receives a plurality of corrected scores CSCRj (j=1 to x×y). The determination unit 320 then performs a "second determination process" to determine whether the first person 1-A and the second person 1-B are the same based on the plurality of corrected scores CSCRj (j=1 to x×y). For example, the determination unit 320 calculates an average score of the plurality of corrected scores CSCRj (j=1 to x×y) and compares the average score with a threshold value TH. If the average score is equal to or greater than the threshold value TH, the determination unit 320 determines that the first person 1-A and the second person 1-B are the same. In this case, the identification information (track ID) of the first track TR-A and the second track TR-B are unified to the same one. On the other hand, if the average score is less than the threshold value TH, the determination unit 320 determines that the first person 1-A and the second person 1-B are not the same.

[0046] The determination section 320 outputs result information 360 indicating the result of the second determination process. The result information 360 can be used for various processes related to the re-identification process.

[0047] 4.Effects As described above, according to this embodiment, the re-identification process is performed taking into consideration the orientation of person 1. More specifically, a rank is assigned to the similarity SIMj in the re-identification process taking into consideration the orientation of person 1. Furthermore, a matching degree αj, which is the degree to which the ranking matches the ranking rule, is calculated. Then, the re-identification process is performed taking into consideration the matching degree αj. In this way, the re-identification process is performed taking into consideration the orientation of person 1, thereby improving the accuracy of the re-identification process.

[0048] The re-identification process according to this embodiment is applicable to both online processing (Multi Track Multi Camera (MTMC)) and offline processing (person retrieval). [Explanation of symbols]

[0049] 1...person, 10...camera, 100...ReID system, 110...processor, 120...storage device, 130...interface, VID...image

Claims

1. A re-identification system for re-identifying a moving object, comprising: one or more processors; one or more storage devices for storing feature information indicating a combination of re-identification features and orientations of moving objects in a video; Equipped with The rank of the similarity between two moving objects is the highest rank when the two moving bodies are identical and the orientations of the two moving bodies are the same; and the lowest rank when the two moving bodies are different and the orientations of the two moving bodies are different; an intermediate rank between the highest rank and the lowest rank; Including, The ranking rule is that the higher the rank, the higher the similarity. The one or more processors: calculating a plurality of similarities between the first moving body and the second moving body based on a plurality of combinations of the re-identified feature amount of the first moving body and the re-identified feature amount of the second moving body; performing a first determination process for provisionally determining whether or not the first moving body and the second moving body are the same based on each of the plurality of similarities; determining a plurality of ranks for each of the plurality of similarities based on a result of the first determination process, the orientation of the first moving body, and the orientation of the second moving body; calculating a degree of correspondence between the plurality of similarities and the plurality of ranks and the ranking rules, the degree being a degree to which the correspondence matches the ranking rules; A second determination process is performed to determine whether the first moving object and the second moving object are the same based on the degree of match and the plurality of similarities. It was configured as Re-identification system.

2. The re-identification system according to claim 1, The one or more processors: calculating a score based on each of the plurality of similarities, the score increasing as each of the similarities increases; performing the first determination process of provisionally determining whether or not the first moving object and the second moving object are the same based on the score; Calculating a corrected score by correcting the score based on the degree of match; The second determination process is performed to determine whether or not the first moving body and the second moving body are the same, based on the corrected score. It was configured as Re-identification system.

3. The re-identification system according to claim 2, a first weighting factor is a weighting factor when the orientation of the first moving body and the orientation of the second moving body are the same; a second weighting factor is a weighting factor when the orientation of the first moving body and the orientation of the second moving body are different, and is smaller than the first weighting factor; The one or more processors: selecting the first weighting factor or the second weighting factor as the weighting factor for each of the plurality of similarities based on the orientation of the first moving body and the orientation of the second moving body; The score is calculated by multiplying each similarity by the weighting factor. It was configured as Re-identification system.

4. The re-identification system according to claim 2, The one or more processors: calculating the degree of match for each of the plurality of similarities by comparing the correspondence between the plurality of similarities and the plurality of ranks with the ranking rule; The score for each of the similarities is multiplied by the degree of match for each of the similarities to correct the score and calculate the corrected score. It was configured as Re-identification system.

5. A re-identification system according to any one of claims 1 to 4, The feature information indicates the combination of the re-identified feature amount and the orientation of the moving object for each track. Re-identification system.

Citation Information

Patent Citations

  • Information processing device and information processing method

    JP2022048817A

  • Object tracking device

    JP2022159694A

  • Device and program for deciding personal action

    JP2011100175A

  • Information processing apparatus, control method therefor, and program

    JP2017033547A

  • Image processing apparatus, method, and program

    JP2017097549A