A pedestrian target association method for surveillance networks based on the maximum slice Wasserstein metric

Abstract

This invention discloses a pedestrian target association method for surveillance networks based on the maximum slice Wasserstein metric. During the recording period, the surveillance network system generates pedestrian target detection and tracking records based on YOLOv5 pedestrian detection and Deepsort, using an n-order dot-mapping mechanism. During the association period, the surveillance network system performs pedestrian target merging and association within a single camera and across cameras on the pedestrian target detection and tracking records formed by all cameras based on the maximum slice Wasserstein distance between n-order continuous high-dimensional feature sequences. After merging and association, based on Openpose human pose detection and Brenner gradient-based image sharpness evaluation, pedestrian target thumbnails are constructed from the video stream, generating records containing pedestrian target ID, n-order continuous high-dimensional feature sequence, camera ID of the pedestrian target, video stream time, and typical thumbnails, which are then stored in a database for querying. This invention provides an effective solution for target association and key target retrieval based on target association in large-scale surveillance networks.

Description

Technical Field

[0001] This invention relates to intelligent analysis of surveillance videos in smart city and smart transportation scenarios, and in particular to a method for pedestrian target association in surveillance networks based on the maximum slice Wasserstein metric. Background Technology

[0002] Public security is a key feature of smart cities, but in recent years, the global security situation has become increasingly complex, with frequent traffic accidents, terrorist attacks, and crimes seriously affecting social stability. To curb and combat crime and reduce social security risks, important locations have installed video surveillance systems to record and store audio and video data. Video surveillance systems are playing a crucial role in early warning of dangerous behavior and in solving cases. my country currently has over 40 million surveillance cameras installed, generating tens of thousands of petabytes (PB) of data annually. The coverage area of ​​video surveillance and the number of monitoring points and checkpoints are increasing rapidly at a rate exceeding 30%. Because video data is unstructured, it is difficult to automatically analyze the rich information it contains. The massive amount of data poses significant technical challenges to the storage, analysis, and retrieval of surveillance videos.

[0003] In recent years, driven by artificial intelligence and big data technologies, the intelligentization of video surveillance systems has developed rapidly, with continuously improving levels of intelligence and automation in task processing. The shift from single-camera monitoring to multi-camera monitoring networks has resulted in a massive increase in video data volume, making more refined and efficient analysis and management based on these networks crucial. Current research on pedestrian target association in single-camera and cross-camera surveillance videos primarily employs separate or integrated approaches to optimize neural networks for pedestrian target feature extraction and target tracking algorithms, aiming to improve the accuracy of target association within the video surveillance network in a short time, with the goal of providing accurate cross-camera trajectories. However, these technical approaches, which focus on improving precise target trajectories, cannot meet the needs of rapid target retrieval and verification over large spatiotemporal ranges.

[0004] In recent years, the Wasserstein metric has been widely used in the field of artificial intelligence. For a metric defined in R... n Given probability distributions μ and ν on the given surface, the Wasserstein distance is used to measure the optimal transport problem from one distribution to another.

[0005]

[0006] Where ξ is an R n ×R n For a joint distribution on the , μ and ν must simultaneously satisfy that μ and ν are its . Marginal distribution d can be R nAny distance on, for example European distance L1 distance, etc.

[0007] Wasserstein was first used in the design of GAN networks. For example, Chinese patent CN113673347A discloses a representation similarity adversarial network based on Wasserstein distance. This prior art solution reduces the marginal probability distribution of source domain subjects and target domain subjects by reducing the Wasserstein distance, and then reduces the conditional probability distribution by association reinforcement. Wasserstein metric is also used for feature similarity calculation. For example, Chinese patent CN110110670B discloses a data association method in pedestrian tracking based on Wasserstein metric. This prior art solution combines Wasserstein distance with neural networks to solve the matching degree of pedestrian features, replacing the traditional Mahalanobis and cosine distances to achieve target tracking.

[0008] While the aforementioned existing technologies are based on Wasserstein metrics for target data tracking, most of them are low-dimensional Wasserstein metrics. High-dimensional Wasserstein metrics are difficult to calculate and are generally calculated by projecting onto a high-dimensional unit sphere to a one-dimensional space to achieve slice calculation. Among all slice calculation results, the largest slice is the one with the largest Wasserstein metric among all projections, but the computational complexity is high. It is especially unsuitable for target tracking and target association in large-scale video surveillance networks. Summary of the Invention

[0009] Purpose of the invention: The purpose of this invention is to provide a pedestrian target association method for surveillance networks based on the maximum slice Wasserstein metric, which provides a target association-based retrieval and query service for large-scale video surveillance networks, thereby improving the retrieval and query efficiency for key targets.

[0010] Technical solution: The present invention provides a method for pedestrian target association in a monitoring network based on the maximum slice Wasserstein metric, comprising the following steps:

[0011] Step S1, during the recording period Q R Within the system, pedestrian target detection is performed on each individual camera in the surveillance network. The pedestrian target detection process uses an n-order dot-mapping mechanism to generate detection and tracking records R for the pedestrian targets. i,j ;

[0012] Step S2, in the recording period Q R The immediately following correlation period Q MWithin the system, the detection and tracking records of pedestrian targets from all cameras in the surveillance network are merged and correlated within each camera. During the merging and correlation process, the maximum slice Wasserstein distance between n-order continuous high-dimensional feature sequences and a predetermined threshold are used as the correlation criteria. After the merging and correlation is completed, a pedestrian target correlation record R' is generated with the camera ID as the primary key. i,j ;

[0013] Step S3: The monitoring network system performs cross-camera merging and association on all pedestrian target records that have already been merged and associated on all cameras. During the merging and association process, the maximum slice Wasserstein distance between n-order continuous high-dimensional feature sequences and a predetermined threshold are used as the association criteria. After the merging and association is completed, a cross-camera pedestrian target association record R” is generated with the pedestrian target ID as the primary key. j ;

[0014] Step S4: The monitoring network system generates records R" based on cross-camera merging and association. j The system includes the camera ID and recording generation time, collects thumbnail images from the surveillance video stream, and evaluates image sharpness based on Openpose human pose detection and Brenner gradient. Then, representative images of the pedestrian's front, back, left, and right sides are selected from the thumbnail set to generate a new record containing the pedestrian's target ID as the primary key and the thumbnail representative images. j And store it in the database for subsequent queries.

[0015] Furthermore, step S1 specifically includes the following steps:

[0016] During the recording period Q R Within the surveillance network system, a single camera C... i This method detects pedestrian targets using YOLOv5 neural network pedestrian detection and Deepsort target tracking algorithm.

[0017] In the aforementioned Deepsort pedestrian target tracking process, an n-order dot-mapping mechanism is continuously used to generate pedestrian target detection and tracking records R. i,j The details are as follows:

[0018] If we are targeting a pedestrian P j The number of high-dimensional pedestrian target features accumulated during continuous tracking in the YOLOv5 output (f' i,j,1 ,f' i,j,2 If the number of features F reaches n, then retain these n features F. i,j =[f i,j,1 ,f i,j,2 ,...,f i,j,nThe system will regenerate a new pedestrian target ID and continue Deepsort tracking. If the number of continuous feature records accumulated during the tracking of a certain pedestrian target is less than n, the detection and tracking record will not be saved.

[0019] Camera C i Every nth order of dot generation generates and outputs the detection tracking record R. i,j ;

[0020] R i,j ={C i P i,j T i,j F i,j =[f i,j,1 ,f i,j,2 ,...,f i,j,n ]}

[0021] Among them, C i Let P be the ID of the i-th camera, where i = 1, 2, ..., N. i,j Let T be the pedestrian target ID of the j-th camera in the i-th camera. i,j To detect and track R i,j The generation time of F i,j It represents n consecutive high-dimensional pedestrian target features output by YOLOv5.

[0022] Furthermore, step S2 specifically includes the following steps:

[0023] S201: In the associated period Q M Internally, the monitoring network system tracks and records all data from all cameras (a total of N cameras) using the R detection data. i,j If we aggregate the data, then camera C i During the recording period Q R The inner constituent set sR i ;sR i ={Ri,1,Ri,2,…};

[0024] S202: For camera C i From centralized sR i The first detection and tracking record R i,1 Begin the merge calculation and calculate R. i,1 Middle feature F i,1 Compared with other detection and tracking records R i,x (e.g. R) i,2 R i,3 The feature F corresponding to (, ...) i,x The maximum slice Wasserstein distance MSW(F) i,1 ,F i,x Then, for all MSW values ​​less than or equal to the predetermined threshold Th1, P...x P1 and P1 belong to the same P' i,1 If there is no P x If it can be merged with P1, then the detection tracking record R will be... i,1 As separate records, the detection tracking record R will be merged. i,x Together with R i,1 From sR i Remove from;

[0025] Repeat the above merge calculation until sR i Set to zero;

[0026] S203: Perform the operation of S202 on all N cameras in the monitoring network, complete the association of pedestrian targets within a single camera, and generate and output a new pedestrian target association record R'. i,j ;

[0027] R' i,j ={C i , P' i,j , F' i,j =(f' i,j,1 ,f' i,j,2 ,...,f' i,j,n ), (T i,j,1 T i,j,2 ...)}

[0028] Among them, F' i,j To merge with the same P' during the association merging process j Other records' high-dimensional feature sequences F i,j The record with the smallest average maximum slice Wasserstein distance corresponds to the F i,j Sequence, (T) i,j,1 T i,j,2 ...) are merged into P' i,j R on all records i,j Recording time T i,j .

[0029] Furthermore, step S3 specifically includes the following steps:

[0030] First, in the associated period Q immediately following the recording period. M Internally, the monitoring network system records R' of all individual camera targets that have been associated and merged. i,j Perform collection, ssR i ={R' i,1 ,R' i,2 ,....};All ssR i Construct a set ssR = {ssR1, ssR2, ..., ssR} N};

[0031] Then, start with the first record R' in ssR1 1,1 Begin cross-camera association merging of pedestrian targets and calculate R' 1,1 China F' 1,1 The recording sets of the other cameras are ssR2, ssR3..., ssR N F' corresponding to each record i,j The maximum slice Wasserstein distance MSW(F) 1,1 ,F i,j Then, for all MSW values ​​less than or equal to the predetermined threshold Th2, P' i,j and P' 1,1 If they are classified into the same P1, then there is no P' i,j Can be used with P' 1,1 Merging will result in P' 1,1 As separate records, merge them into P”1; and group R’ 1,1 Remove from sR1, and associate with R' i,j From sR i Remove from;

[0032] Repeat the above merge calculation in sR1 until the sR1 set is cleared to zero;

[0033] After clearing sR1, then process sR2,...,sR in sequence. N Repeat the above merge calculation until a subset ssR in ssR is reached. i All records in the database are cleared;

[0034] Finally, the monitoring network system generates cross-camera pedestrian target association records R”. j ;

[0035] R” j ={P” j ,F” j =(f” i,j,1 ,f” i,j,2 ,...,f” i,j,n ), [(C i1 T i1,j1 T i,j2 ,...), (C i2 T i2,j1 T i2,j2 ,...),....]}

[0036] Where F” j F' is the record with the smallest Wasserstein distance to the average maximum slice of all other records belonging to the same objective. i,j , (C i1 T i1,j1 Ti1,j2 ...) is the pedestrian target P” j Appearing in camera C i1 The recorded times T i1,j1 T i,j2 ,...

[0037] Furthermore, step S4 specifically includes the following steps:

[0038] S401: The monitoring network system associates pedestrian targets across the camera with "R". j According to (C) i1 T i1,j1 T i,j2 ,...), (C i2 T i2,j1 T i2,j2 ...), from the corresponding camera C in the surveillance network ix The moment T ix,jy The video stream captures m images at intervals, forming a sequence with R”. j Corresponding thumbnail atlas T; Camera C ix Specifically, it refers to the camera number where the pedestrian is located when the pedestrian is spotted;

[0039] S402: Based on the Openpose human pose detection algorithm, the thumbnail set T is classified into T1, T2, T3, T4, T5, T6, T7, T8, T9, T1, T1, T2, T3, T4, T1, T2, T3, T4, T1, T2, T3, T4, T4, T5, T6, T7, T8, T9, T1, T1, T2, T3, T4 ... f T b T l and T r ;

[0040] S403: An image sharpness evaluation algorithm based on Brenner gradient, which evaluates sharpness from T... f T b T l and T r Select the representative image with the highest clarity from the image. f I b I l and I r ;

[0041] S404: Monitoring network system generates new record R”' j ={P” j 、(I f I b I s ), F” j =(f” i,j,1 ,f” i,j,2 ,...,f” i,j,n ), [(C i1 T i1,j1 T i,j2,...), (C i2 T i2,j1 T i2,j2 ,...),....]}, and record R”' j Stored in the database for later retrieval.

[0042] The present invention also discloses a computer storage medium storing a computer program that, when executed by a processor, implements the above-described method for pedestrian target association in a monitoring network based on the maximum slice Wasserstein metric.

[0043] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a pedestrian target association method for a monitoring network based on the maximum slice Wasserstein metric.

[0044] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0045] 1. Unlike traditional single-camera and cross-camera pedestrian re-identification technologies that operate within a short period of time, this invention enables rapid retrieval of relevant video streams of important target objects in video surveillance networks;

[0046] 2. This invention uses an n-order dotting mechanism to retain n consecutive high-dimensional features of the tracked target, and uses the more effective maximum slice Wasserstein distance between the statistical distribution of the high-dimensional vector space of the feature set of the pedestrian target at different angles implied by the n-order high-dimensional features as the basis for similarity measurement. It is superior to the cosine distance or Euclidean distance of the single-frame features used by the mainstream in terms of multi-order feature design and similarity measurement specifications.

[0047] 3. This invention adopts an alternating recording period and association period to alleviate the impact of factors such as changes in pedestrians' clothing when performing target association over a large time span, thus balancing the accuracy of association and the efficiency of actual use. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0049] Figure 2 This is a flowchart of an embodiment;

[0050] Figure 3 This is a schematic diagram illustrating the merging of detection records in an embodiment. Detailed Implementation

[0051] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0052] like Figure 1 As shown, this invention provides a pedestrian target association method for monitoring networks based on the maximum slice Wasserstein metric, within the recording period Q. R Within the system, the monitoring network system uses YOLOv5 neural network for pedestrian detection and Deepsort, employing an n-order dot-mapping mechanism to generate detection and tracking records for pedestrian targets; during the correlation period Q... M Within the system, the monitoring network performs pedestrian target merging and association within a single camera and across cameras based on the maximum slice Wasserstein distance between n-order continuous high-dimensional feature sequences and the detection and tracking records of pedestrian targets formed by all cameras. After merging and association, pedestrian target thumbnails are constructed from the video stream based on Openpose human pose detection and Brenner gradient-based image sharpness evaluation. Records containing pedestrian target ID, n-order continuous high-dimensional feature sequence, camera ID of the pedestrian target, video stream time, and typical thumbnails are generated and stored in the database for querying.

[0053] The recording period Q of the rolling loop of this invention R and related period Q M Based on single-frame YOLOv5 target detection and continuous multi-frame Deepsort tracking algorithms, this paper adopts an n-order dotting mechanism to divide the target tracking data into equal segments; and realizes target association within a single camera and between multiple cameras through a multi-order high-dimensional feature similarity comparison algorithm based on the maximum slice Wasserstein metric. Ultimately, it can realize target association in a large-scale monitoring network and key target retrieval based on target association.

[0054] The specific steps in this embodiment are as follows:

[0055] S1: During the recording period Q R Within the surveillance network system, each individual camera is based on the existing YOLOv5 pedestrian detection and Deepsort target tracking algorithms for pedestrian target detection. The key point of this embodiment is that an n-order dot-mapping mechanism is continuously used to generate pedestrian target detection and tracking records during the pedestrian target detection process.

[0056] S2: In the associated period Q immediately following the recording period M In this embodiment, the recording period is set to nighttime. The monitoring network system performs merging and association of the pedestrian target detection and tracking records formed by all cameras within a single camera. During the merging and association process, the maximum slice Wasserstein distance between n-order continuous high-dimensional feature sequences and a predetermined threshold are used as the association basis. After merging, pedestrian target association records are generated with camera ID as the primary key.

[0057] S3: After completing the pedestrian target association and merging for a single camera, the monitoring network system performs cross-camera merging and association on the target records that have already been merged on all cameras. During the merging and association process, the maximum slice Wasserstein distance between n-order continuous high-dimensional feature sequences and the predetermined threshold are still used as the association basis. After merging, pedestrian target association records are generated with pedestrian target ID as the primary key.

[0058] S4: The monitoring network system collects thumbnail images from the monitoring video stream based on the camera ID and record generation time contained in the records generated by cross-camera merging and association. Based on OpenPose human pose detection and Brenner gradient-based image sharpness evaluation, it selects representative images of the front, back, left, and right sides of pedestrians from the thumbnail image set, generates new records containing the representative thumbnail images with the pedestrian target ID as the primary key, and stores them in the database for subsequent queries.

[0059] In this embodiment, step S1 specifically includes the following steps:

[0060] During the recording period Q R In this embodiment, the recording period is set to daytime, and each camera in the monitoring network system detects pedestrian targets based on the YOLOv5 neural network pedestrian detection and the Deepsort target tracking algorithm;

[0061] Camera C i During the Deepsort pedestrian target tracking process described above, an n-order point-marking mechanism is continuously employed (for surveillance videos with a frame rate of 30, typically 5 consecutive seconds are taken, i.e., n=150), as detailed below:

[0062] For a single target, if for target P j The number of high-dimensional pedestrian target features accumulated during continuous tracking by YOLOv5 output (f') i,j,1 ,f' i,j,2 If the number of features F reaches n, then retain n features F. i,j =[f i,j,1 ,f i,j,2 ,...,f i,j,n The new target ID is regenerated and Deepsort tracking continues; if the number of consecutive feature records accumulated for the tracked object is less than n, the record is not saved.

[0063] Ultimately, camera C i Generate and output detection tracking records R i,j ={C i P i,j T i,j F i,j =[f i,j,1 ,f i,j,2,...,f i,j,n ]}, where C i For camera ID, P i,j For pedestrian ID, T i,j For R i,j Record generation time, F i,j It represents n consecutive high-dimensional pedestrian target features output by YOLOv5.

[0064] In this embodiment, step S2 specifically includes the following steps:

[0065] S201: In the associated period Q immediately following the recording period M In this embodiment, the recording period is set to nighttime, and then all recordings R from a total of N cameras are recorded. i,j Collect data and record camera C. i During the recording period Q R The inner constituent set sR i ;

[0066] S202: For camera C i From sR i The first record R i,1 Begin the merge calculation and calculate R. i,1 China F i,1 With the rest of the records R i,2 R i,3 , ..., corresponding F i,x The maximum slice Wasserstein distance MSW(F) i,1 ,F i,x ), and P values ​​less than or equal to a predetermined threshold Th1. x P1 and P1 belong to the same P' i,1 If there is no P x If it can be merged with P1, then R will be... i,1 Merge records as independent records, and record R i,x Together with R i,1 From sR i Remove from the middle; repeat the above merge calculation until sR i Set to zero.

[0067] S203: Implement the operation of S202 on all N cameras in the monitoring network, complete the association of pedestrian targets within a single camera, and generate and output a new record R'. i,j ={C i , P' i,j , F' i,j =(f' i,j,1 ,f' i,j,2 ,...,f' i,j,n ), (T i,j,1 T i,j,2...)}, where F' i,j To merge with the same P' during the association merging process j Other records' high-dimensional feature sequences F i,j The record with the smallest average maximum slice Wasserstein distance corresponds to the F i,j Sequence, (T) i,j,1 T i,j,2 ...) are merged into P' i,j R on all records i,j Recording time T i,j .

[0068] In this embodiment, step S3 specifically includes the following steps:

[0069] S301: In the associated period Q immediately following the recording period M Within this embodiment, the recording period is set to nighttime. The monitoring network system records R' of all single-channel camera targets that have been associated and merged by each camera. i,j Perform collection, ssR i ={R' i,1 ,R' i,2 ,....};Record all ssR i Construct a set ssR = {ssR1, ssR2, ..., ssR} N}

[0070] S302: Start from the first record R' of ssR1 1,1 Begin the association and merging of the objectives, and calculate R'. 1,1 China F' 1,1 The recording sets of the other cameras are ssR2, ssR3..., ssR N F' corresponding to each record i,j The maximum slice Wasserstein distance MSW(F) 1,1 ,F i,j ), and P' for all MSW values ​​less than or equal to the predetermined threshold Th2. i,j and P' 1,1 If they are classified into the same P1, then there is no P' i,j Can be used with P' 1,1 Merging will result in P' 1,1 As separate records, merge them into P”1; and group R’ 1,1 Remove from sR1, and associate with R' i,j From sR i Remove from the middle; repeat the above merge calculation in sR1 until the sR1 set is zeroed; after sR1 is zeroed, perform the merge calculation on sR2,...,sR... N Repeat the above merge calculation until a subset ssR in ssR is reached.i All records in the database were cleared.

[0071] S303: After completing step S302, the monitoring network system generates a new record R. j ={P” j ,F” j =(f” i,j,1 ,f” i,j,2 ,...,f” i,j,n ), [(C i1 T i1,j1 T i,j2 ,...), (C i2 T i2,j1 T i2,j2 ,...),....]},where F” j F' is the record with the smallest Wasserstein distance to the average maximum slice of all other records belonging to the same objective. i,j , (C i1 T i1,j1 T i1,j2 ...) is the target P” j Appearing in camera C i1 The recorded times T i1,j1 T i,j2 ,...

[0072] In this embodiment, step S4 specifically includes the following steps:

[0073] S401: The monitoring network system, according to R” j According to (C) i1 T i1,j1 T i,j2 ,...), (C i2 T i2,j1 T i2,j2 ...), from the corresponding camera C in the surveillance network ix The moment T ix,jy The video stream captures m images at intervals, forming a sequence with R”. j The corresponding thumbnail set T;

[0074] S402: Based on the OpenPose human pose detection algorithm, the thumbnail set T is classified into T1, T2, T3, T4, T5, T6, T7, T8, T9, T1, T1, T2, T3, T4 ...3, T4, T4, T5, T6, T7, T8, T9, T1, T1, T2, T3, f T b T l and T r ;

[0075] S403: An image sharpness evaluation algorithm based on Brenner gradient, which evaluates sharpness from T... f T bT l and T r Select the representative image with the highest clarity from the image. f I b I l and I r ;

[0076] S404: Monitoring network system generates new record R”' j ={P” j 、(I f I b I s ), F” j =(f” i,j,1 ,f” i,j,2 ,...,f” i,j,n ), [(C i1 T i1,j1 T i,j2 ,...), (C i2 T i2,j1 T i2,j2 , ...), ....]} and store them in the database for querying.

[0077] Example 1:

[0078] like Figure 2 and Figure 3 As shown, the pedestrian targets in the surveillance video database of this embodiment are tracked and associated, and the specific content is as follows:

[0079] For images from N cameras in the surveillance video database, first apply the following to each camera sequentially: Figure 2 Pedestrian detection using YOLOv5 neural network and Deepsort target tracking algorithm are performed to identify pedestrian targets and generate detection and tracking records R. i,j Then they are grouped into sR1, sR2, ..., sR N Next, the records of all single-channel camera targets that have been associated and merged among these N cameras are set to obtain ssR, and the ssR set is further processed by the set of ssR. i The data are merged separately to generate the target trajectory database R. j (See Figure 2 and Figure 3 ).

Claims

1. A method for pedestrian target association in a surveillance network based on the maximum slice Wasserstein metric, characterized in that, Includes the following steps: Step S1, during the recording period Q R Within the system, pedestrian target detection is performed on each individual camera in the surveillance network. The pedestrian target detection process uses an n-order dot-mapping mechanism to generate detection and tracking records R for the pedestrian targets. i,j ; Step S2, in the recording period Q R The immediately following correlation period Q M Within the system, the detection and tracking records of pedestrian targets from all cameras in the surveillance network are merged and correlated within each camera. During the merging and correlation process, the maximum slice Wasserstein distance between n-order continuous high-dimensional feature sequences and a predetermined threshold are used as the correlation criteria. After the merging and correlation is completed, a pedestrian target correlation record R' is generated with the camera ID as the primary key. i,j ; Step S3: Perform cross-camera merging and association on all pedestrian target records that have already been merged and associated on all cameras in the monitoring network system. During the merging and association process, the maximum slice Wasserstein distance between n-order continuous high-dimensional feature sequences and a predetermined threshold are used as the association criteria. After the merging and association is completed, a cross-camera pedestrian target association record R'' is generated with the pedestrian target ID as the primary key. j ; Step S4: Based on the cross-camera pedestrian target association record R'' j The dataset includes camera IDs and recording times, and thumbnail sets T are collected from the surveillance video stream. Image sharpness is evaluated based on Openpose human pose detection and Brenner gradient. Next, representative images of the pedestrian's front, back, left, and right sides are selected from the thumbnail set, generating a new record R''' containing the representative thumbnail images with the pedestrian's target ID as the primary key. j And store it in the database for subsequent queries.

2. The pedestrian target association method for monitoring networks based on the maximum slice Wasserstein metric as described in claim 1, characterized in that, Step S1 specifically includes the following steps: During the recording period Q R Inside, for a single camera C in the monitoring network system i Pedestrian target detection is performed based on YOLOv5 neural network pedestrian detection and Deepsort target tracking algorithm; Camera C i The Deepsort pedestrian target tracking process employs an n-order point-marking mechanism, specifically as follows: If a pedestrian target P is being tracked... j The number of high-dimensional pedestrian target features accumulated during continuous tracking by YOLOv5 output (f' i,j,1 ,f' i,j,2 If the number of features F reaches n, then retain these n features F. i,j =[f i,j,1 , f i,j,2 ,..., f i,j,n The system will regenerate a new pedestrian target ID and continue Deepsort tracking. If the number of continuous feature records accumulated during the tracking of a certain pedestrian target is less than n, the detection and tracking record will not be saved. Camera C i Every nth order of dot generation generates and outputs the detection tracking record R. i,j ; R i,j ={C i ,P i,j ,T i,j ,F i,j =[f i,j,1 , f i,j,2 ,..., f i,j,n ]} Among them, C i Let P be the ID of the i-th camera, i = 1, 2, ..., N. i,j Let T be the pedestrian target ID of the j-th camera in the i-th camera. i,j To detect and track R i,j The generation time of F i,j It represents n consecutive high-dimensional pedestrian target features output by YOLOv5.

3. The pedestrian target association method for monitoring networks based on the maximum slice Wasserstein metric as described in claim 1, characterized in that, Step S2 specifically includes the following steps: S201: In the associated period Q M Within, all detection and tracking records R from N cameras i,j If we aggregate the data, then camera C i During the recording period Q R The inner constituent set sR i sR i ={Ri,1,Ri,2,…}; S202: For camera C i From centralized sR i The first detection and tracking record R i,1 Begin merging and calculating the detection tracking record R. i,1 Middle feature F i,1 Compared with other detection and tracking records R i,x The corresponding feature F i,x The maximum slice Wasserstein distance MSW(F) i,1 , F i,x Then, for all MSW values ​​less than or equal to the predetermined threshold Th1, P... x P1 and P1 are grouped into the same P' i,1 If there is no P x If it can be merged with P1, then the detection tracking record R will be... i,1 As separate records, the detection tracking record R will be merged. i,x Together with R i,1 From sR i Remove from middle; R i,x ={R i,2 R i,3 , ...}; Repeat the above merge calculation until sR i Set to zero; S203: Merge all cameras separately, complete the pedestrian target association within each camera, and generate and output a new pedestrian target association record R'. i,j ; R' i,j ={C i ,P' i,j ,F' i,j =(f' i,j,1 , f' i,j,2 ,..., f' i,j,n ),(T i,j,1 、T i,j,2 、...)} Among them, F' i,j To merge with the same P' during the association merging process j Other records' high-dimensional feature sequences F i,j The record with the smallest average maximum slice Wasserstein distance corresponds to the F i,j Sequence, (T) i,j,1 T i,j,2 ...) are merged into P' i,j R on all records i,j Recording time T i,j .

4. The pedestrian target association method for monitoring networks based on the maximum slice Wasserstein metric as described in claim 3, characterized in that, Step S3 specifically includes the following steps: First, in the associated period Q immediately following the recording period M Within, the records R' of all single-channel camera targets that have been associated and merged for each camera. i,j Perform collection, ssR i ={R' i,1 ,R' i,2 ,....};All ssR i Construct a set ssR = {ssR1, ssR2, ..., ssR} N }; Then, start with the first record R' in ssR1 1,1 Begin cross-camera association merging of pedestrian targets and calculate R' 1,1 China F' 1,1 The recording sets of the other cameras are ssR2, ssR3..., ssR N F' corresponding to each record i,j The maximum slice Wasserstein distance MSW(F) 1,1 , F i,j Then, for all MSW values ​​less than or equal to the predetermined threshold Th2, P' i,j and P' 1,1 If they are grouped into the same P''1, then there is no P' i,j Can be used with P' 1,1 Merging will result in P' 1,1 As separate records, merge them into P''1; and group R''... 1,1 Remove from sR1, and associate with R' i,j From sR i Remove from; Repeat the above merge calculation in sR1 until the sR1 set is cleared to zero; After clearing sR1, then process sR2,...,sR in sequence. N Repeat the above merge calculation until a subset ssR in ssR is reached. i All records in the database are cleared; Finally, a cross-camera pedestrian target association record R'' is generated. j ; R'' j ={P'' j ,F'' j =(f'' i,j,1 , f'' i,j,2 ,..., f'' i,j,n ),[(C i1 ,T i1,j1 ,T i,j2 ,...),(C i2 ,T i2,j1 ,T i2,j2 ,...),....]} Where F'' j F' is the record with the smallest Wasserstein distance to the average maximum slice of all other records belonging to the same objective. i,j , (C i1 T i1,j1 T i1,j2 ...) represents the pedestrian target P'' j Appearing in camera C i1 The recorded times T i1,j1 T i,j2 ,...

5. The pedestrian target association method for monitoring networks based on the maximum slice Wasserstein metric according to claim 1, characterized in that, Step S4 specifically includes the following steps: S401: Associating R'' with cross-camera pedestrian targets j According to (C) i1 T i1,j1 T i,j2 ,...), (C i2 T i2,j1 T i2,j2 ...), obtain the corresponding camera C from the monitoring network. ix T ix,jy In the real-time video stream, m images are captured at intervals, forming a sequence with R'' j The corresponding thumbnail set T; S402: Based on the Openpose human pose detection algorithm, the thumbnail set T is classified into T1, T2, T3, T4, T5, T6, T7, T8, T9, T1, T1, T2, T3, T4 ...4, T5, T6, T7, T8, T9, T1, T1, T2, T3, T4, T4, T f T b T l and T r ; S403: An image sharpness evaluation algorithm based on Brenner gradient, which evaluates sharpness from T... f T b T l and T r Select the representative image with the highest clarity from the image. f I b I l and I r ; S404: Monitoring network system generates new record R''' j ={P'' j 、(I f I b I s ), F'' j =(f'' i,j,1 , f'' i,j,2 ,..., f'' i,j,n ), [(C i1 T i1,j1 T i,j2 ,...), (C i2 T i2,j1 T i2,j2 ,...),....]}, and record R''' j Stored in the database for later retrieval.

6. A computer storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the pedestrian target association method for a monitoring network based on the maximum slice Wasserstein metric as described in any one of claims 1-5.

7. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the pedestrian target association method for monitoring networks based on the maximum slice Wasserstein metric as described in any one of claims 1-5.

Images