CN112131929B - Cross-camera pedestrian tracking system and method based on block chain - Google Patents
Cross-camera pedestrian tracking system and method based on block chain Download PDFInfo
- Publication number
- CN112131929B CN112131929B CN202010783800.0A CN202010783800A CN112131929B CN 112131929 B CN112131929 B CN 112131929B CN 202010783800 A CN202010783800 A CN 202010783800A CN 112131929 B CN112131929 B CN 112131929B
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- tracked
- module
- short
- short track
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N7/00—Television systems
- H04N7/18—Closed-circuit television [CCTV] systems, i.e. systems in which the video signal is not broadcast
- H04N7/181—Closed-circuit television [CCTV] systems, i.e. systems in which the video signal is not broadcast for receiving images from a plurality of remote sources
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Signal Processing (AREA)
- Human Computer Interaction (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to a cross-camera pedestrian tracking system and a cross-camera pedestrian tracking method based on a block chain, which comprise the following 6 modules: the system comprises an entrance module, a data collection module, a short track generation module, a tracking module, an output module and a block chain sharing module. The invention mainly completes the pedestrian tracking task and function deployed at the camera end under the cross-camera scene. The user can utilize the system to complete the tracking of the pedestrian under the cross-camera monitoring scene and keep the consistency of the pedestrian label under the cross-camera monitoring scene.
Description
Technical Field
The invention relates to a block chain-based cross-camera pedestrian tracking system and a block chain-based cross-camera pedestrian tracking method, in particular to a near real-time pedestrian tracking system which uses a block chain technology to complete cross-camera pedestrian tracking information sharing, does not depend on a central server, can be deployed to a camera end, provides historical movement tracks of pedestrians in a cross-camera scene, and belongs to the field of monitoring and security.
Background
Currently, a surveillance video acquired by a surveillance camera is generally uploaded to a central server, and the server performs video analysis tasks, such as pedestrian detection, pedestrian tracking, and the like. However, in practical applications, due to the limitations of various factors such as the speed of the network, the capacity of the server, and the performance of the server, the task of performing pedestrian tracking on the central server in a centralized manner is very difficult, and the requirement of near-real-time pedestrian tracking cannot be met.
In a traditional pedestrian tracking method, monitoring videos collected by a plurality of cameras are transmitted to a central server in a centralized manner. And the monitoring videos are processed in a centralized manner on the central server, and then the pedestrian tracking results of the monitoring videos are further fused. The main flow frame is based on detection tracking, namely a detection frame of the pedestrian is firstly obtained, then the appearance characteristic of the pedestrian is extracted, the frame is associated by utilizing the appearance information and the frame position, and finally a complete pedestrian track is formed. If the pedestrian tracking task across the cameras needs to be processed, the pedestrian track obtained by a single camera is further processed to obtain the pedestrian track under the cameras. Currently, the main problems in cross-camera pedestrian tracking research are the processing speed of the tracking algorithm and the credibility of the cross-camera pedestrian tracking information.
At present, pedestrian tracking algorithms have achieved a lot of research results, but there are some difficulties in speed and resource consumption, and pedestrian tracking across cameras remains a very challenging problem. The method has the advantages that the method not only has some common classical problems in the field of computer vision, but also has the inherent defects of detection by deep learning technology. How to achieve real-time or near real-time pedestrian tracking in an environment with limited computing resources of edge computing devices such as cameras is a significant problem in the field of security. The efficient pedestrian tracking method and the credible information sharing mode can greatly improve the effect of cross-camera pedestrian tracking, and in a time monitoring scene, complex shielding puts very severe requirements on a pedestrian tracking algorithm, so that the research of the anti-shielding cross-camera pedestrian tracking algorithm is still one of the centers of gravity of the current research.
In addition, most of the current pedestrian tracking applications use a centralized processing framework for processing, and although this framework can complete most of the pedestrian tracking tasks, the required computational resources and time consumption are very large, and are not acceptable in many delay-sensitive applications. How to increase the speed of the tracking algorithm is also one of the development directions of the core problems in the field of computer vision at present.
The invention aims at the near real-time pedestrian tracking requirement with high accuracy so as to improve the capability of the monitoring system in responding to the flow direction tracking and public opinion control of the crowd in multiple scenes. By utilizing the capacity of the block chain for data sharing, an efficient data association method is established, and the near-real-time cross-camera tracking requirement can be met.
In the aspect of scientific research, the system meets the research requirements of various tracking vision tasks, provides a near-real-time pedestrian tracking system with high accuracy and low energy consumption, and provides substantial help for urban intelligent management.
Disclosure of Invention
The invention solves the problems: the defects in the prior art are overcome, the cross-camera pedestrian tracking system based on the block chain is provided, the cross-camera pedestrian tracking information interaction is completed by using the block chain technology, and an efficient high-shielding tracking method is designed.
The invention adopts the following technical scheme:
the invention relates to a pedestrian tracking system across cameras based on a block chain, which comprises: the system comprises an entrance module, a data collection module, a short track generation module, a tracking module, an output module and a block chain sharing module; wherein:
the entrance module provides a management interface, so that a user can conveniently manage the system from a graphical terminal, and the user selects to enter the data collection module, the short track generation module, the tracking module, the output module and the block chain sharing module according to the actual requirement of the user; the user login interface is used, a user name and a password are input when the user logs in, and corresponding permission is opened by checking the user name and the password; for a common user, only online viewing permission is provided, and various data cannot be exported; for the administrator, providing all the authorities and exporting various data from the data export interface provided by the entrance module;
the data collection module is responsible for collecting data, including real-time video data (24 fps) of the camera, and acquiring the number of the current camera; shooting images of a monitored scene through a camera, and acquiring continuous video image frames in a monitored space within 24 hours from a lead-out interface provided by the camera; transmitting the serial number of the camera to a block chain module, and transmitting the video image frame to a short track generation module;
short track generation module: firstly, receiving video image frames transmitted by a data collection module, and processing the 8 video image frames by using a YOLOV3 detection network every time 8 video image frames are received to obtain pedestrian detection frames in the 8 video image frames; according to the obtained detection frames, cutting corresponding regions from 8 video frames to form pedestrian cutting regions, and packaging the pedestrian cutting regions and the corresponding detection frames of pedestrians into detection sheet images of the pedestrians; clustering the pedestrian detection sheet image by using a clustering method to generate a short track with the minimum length of 3 frames and the maximum length of 8 frames, extracting the characteristics of the short track by using a co-occurrence constraint network after training, and finally transmitting the short track and the extracted characteristics of the short track into a tracking module;
the system comprises a tracking module, a block chain-based cross-camera pedestrian tracking method, a plurality of pedestrian tracking module and a database, wherein the plurality of data are maintained for each pedestrian, and part of the data are stored in the database, the plurality of data comprise a feature pool for storing the appearance features of the pedestrian, the number of times of miss, a historical track, the current state of the pedestrian, the serial number of the pedestrian, the number of times of hit and the current position of the pedestrian, the number of times of hit refers to the number of times that the pedestrian successfully obtains short track matching, and the number of times of miss refers to the number of frames that the pedestrian successfully obtains matching distance from the current frame last time; the current state comprises one of sensitivity, confirmation and deletion (wherein sensitivity means that the pedestrian is possibly caused by false detection, confirmation means that the pedestrian is being tracked, and deletion means that the pedestrian is lost from the detection area); when the module is implemented, firstly, the short track generated by the short track generation module and the characteristics of the short track are received, then the state of the currently tracked pedestrian is updated, and a similarity matrix between the currently tracked pedestrian and the received short track is calculated; then according to the position of the current tracked pedestrian and the input short track, filtering the similarity matrix to obtain a filtered similarity matrix; the optimal distribution scheme of the filtered similarity matrix is solved by using a Hungarian algorithm, a plurality of paired matrix subscripts are obtained, namely the successfully matched pedestrians and short tracks are obtained, and the unsuccessfully matched pedestrians and short tracks which are not obtained are called unsuccessfully matched pedestrians; adding one to the number of times of miss of the unsuccessfully matched pedestrians, and initializing unsuccessfully matched short tracks as new pedestrians; after all unsuccessfully matched pedestrians and short tracks are processed, updating a feature pool of the pedestrian appearance features of which the current states are confirmed and tracked by using a feature restoration algorithm based on time simulated annealing; after the feature pools of the appearance features of all tracked pedestrians are updated, adding all tracked pedestrians by using a confidence-based virtual short track algorithm, namely, estimating the short tracks according to historical tracks of the pedestrians instead of being generated by a short track generation module; adding the virtual short trajectory to a historical trajectory of the tracked pedestrian; if the entrance module finishes the tracking command, storing the historical tracks of all tracked pedestrians in a database;
the block chain module is responsible for sharing the data of the pedestrians tracked among different cameras; the module detects the state change of the tracking module for tracking the pedestrian, if the state of one tracked pedestrian in the tracking module is changed into deletion, the characteristic pool and the serial number of the pedestrian as well as the serial number of the current camera are packed into a block, and the block is sent to a block chain; the module continuously detects whether a new block exists in the block chain or not, when the new block is uploaded to the block chain, data in the new block are downloaded, if the serial number of the camera stored in the new block is different from that of the current camera, serial numbers of a feature pool and pedestrians in the new block are downloaded, a ReID (people re-identification) network is used for carrying out people re-identification on all pedestrians tracked by the feature pool and the current tracking module, and if the ReID network judges that the feature pool is consistent with a certain tracked pedestrian, a fusion serial number request is sent to the tracking module, and the request comprises the serial number downloaded from the new block and the corresponding serial number of the pedestrian tracked by the current tracking module;
the output module provides two data interfaces, one data interface only provides the view of the historical track of the pedestrian in the database, and the other data interface not only provides the view, but also provides an interface for deriving the historical track of the pedestrian; the module provides an interface for a user to view information, and encapsulates the rest of the modules, so that each module is transparent to the user.
The invention has the innovation points that a co-occurrence constraint network, a pedestrian tracking method based on a block chain, a characteristic repairing algorithm based on time simulation annealing and a virtual short track algorithm based on confidence coefficient are adopted, wherein the co-occurrence constraint network is used for a short track generation module to extract short track characteristics, the cross-camera pedestrian tracking method based on the block chain is realized by a tracking module and the block chain module, and the characteristic repairing algorithm based on the time simulation annealing and the virtual short track algorithm based on the confidence coefficient are realized in the tracking module.
In the short track generation module, the structure and training process of the co-occurrence constraint network are as follows:
(1) The co-occurrence constraint network receives a short track with the length of 8 as input, and if the given short track is less than 8, the frame is interpolated by a linear interpolation method to ensure that the length of the short track is 8;
(2) As shown in fig. 2, a co-occurrence constrained network is established, a network structure before a ResNet50 global average pooling layer is used as a backbone network, a fixed division layer is used for uniformly dividing a feature map output by the backbone network, and the feature map is divided into 6 blocks, wherein each block is called a co-occurrence local feature; processing all co-occurrence local features by using a 1-by-1 convolution kernel and an average pooling layer, and reducing the dimensionality of the co-occurrence local features to obtain compressed features; the compressed features are input into a multi-head attention layer to calculate spatial co-occurrence constraints among the compressed features; meanwhile, a plurality of compression features are spliced into individual features; the structure from the backbone network to the completion of the splicing of the compression features is shared by all frames in the short track, that is, each frame in the short track is subjected to the same structural processing, and finally, respective individual features and respective spatial co-occurrence constraints are obtained; the spatial co-occurrence constraints of all frames are packed into a sequence and input into an LSTM layer of 256 hidden units, after the output of the LSTM layer passes through a random discarding layer, the output of the random discarding layer is input into an LSTM layer of another 256 hidden units, the output of the LSTM layer is input into another random discarding layer, the output of the random discarding layer is input into a full connection layer, the full connection output is converted into a group of weights by a softmax layer, and the individual features of all frames are combined into short track features by using the reorganization time weights; the short track characteristics are input into a full connection layer to carry out short track classification, namely, the last full connection layer of the network outputs a label corresponding to the short track;
(3) The co-occurrence constraint network completes training on an iLIDS-VID (public data set on the Internet), video data in the data set is segmented into short tracks with the length of 8, and the short tracks generated by the same video have the same label; all the short tracks are input into a co-occurrence constraint network, and the short tracks are classified to obtain predicted labels; updating weight parameters of all layers behind a backbone network of the co-occurrence constrained network by a gradient descent method by calculating loss of labels predicted by the co-occurrence constrained network and real labels in a data set; after loss convergence, fixing and storing the weight parameters, and simultaneously storing the network structure of the co-occurrence constraint network;
(4) And the short track generation module loads weight parameters of the co-occurrence constraint network after pre-training, and performs feature extraction on the short tracks generated by clustering by using the network to obtain features of the short tracks.
In the tracking module, a cross-camera pedestrian tracking method based on a block chain is as follows:
(1) Receiving the short track and the feature of the short track transmitted by the short track generation module, wherein the feature is a feature vector with 1536 dimensions;
(2) If the currently tracked pedestrian is empty, turning to the step (6); otherwise, the current status of all tracked pedestrians is updated: if the current state of the tracked pedestrian is sensitive and the hit frequency is more than or equal to 2, changing the state of the tracked pedestrian into confirmation; if the current state of the tracked pedestrian is confirmation and the number of times of miss is 18 or more, changing the state thereof to deletion; if the current state of the tracked pedestrian is sensitive, the number of times of miss is more than or equal to 3, and the number of times of hit is equal to 0, the state of the tracked pedestrian is changed into deletion;
(3) Establishing a similarity matrix, calculating the cosine distance between the features in the feature pool of the currently tracked pedestrian and the features of the short track, and establishing the similarity matrix by taking the obtained cosine distance as a basic element;
(4) Filtering a similarity matrix, calculating the distance between the position of the current tracked pedestrian and the position of the short track, and setting the cosine distance between the tracked pedestrian with the distance larger than a threshold value and the short track in the similarity matrix as 0;
(5) Distributing, wherein the optimal solution of the similarity matrix is a bipartite graph problem, the solution is carried out by using a Hungarian algorithm, a plurality of successfully tracked pedestrians and short tracks are obtained for matching, and the rest are unsuccessfully matched short tracks and unsuccessfully matched tracked pedestrians;
(6) Initializing a new tracked pedestrian, initializing unsuccessfully matched short tracks as the new tracked pedestrian, wherein the state of the new tracked pedestrian is sensitive, adding appearance features to the tracks into a feature pool of the new tracked pedestrian, the position of the new tracked pedestrian is the average value of frames of the short tracks, the miss times and hit times of all the frames of the historical tracks, which are the short tracks, are initialized to be 0, and the serial number is the maximum serial number of the currently tracked pedestrian plus 1;
(7) Updating the feature pool of the tracked pedestrian by using a feature repairing algorithm based on time simulated annealing, repairing the appearance feature of the short track matched with the confirmed tracked pedestrian by using the algorithm to obtain a repaired feature, and adding the repaired feature into the confirmed tracked pedestrian feature pool;
(8) Adding virtual short tracks to all tracked pedestrians by using a confidence coefficient-based virtual short track algorithm, and adding the virtual short tracks to historical tracks of the tracked pedestrians;
(9) Detecting whether a block chain module has a request for fusing serial numbers, and if so, modifying the serial number of the currently tracked pedestrian according to the serial number contained in the request; if no, turning to the step (10);
(10) If a command of ending the tracking is received, filtering the tracked pedestrians according to the ratio of the number of the virtual short tracks to the number of the real tracks, deleting the pedestrians higher than a threshold value, storing the remaining pedestrians in a database, and ending; and (4) if the command of ending the tracking sent by the entrance module is not received, the step (1) is carried out.
In the tracking module, a characteristic repairing algorithm based on time simulated annealing is as follows:
(1) The input of the characteristic repairing algorithm based on time simulated annealing is the current tracked pedestrian, the short track transmitted by the short track generating module, and the matching between the successfully tracked pedestrian and the short track obtained by the pedestrian tracking method based on the block chain, which is called as a matching set hereinafter;
(2) If the matching set is empty, the step (7) is carried out; otherwise, taking out a match from the matching set in sequence, wherein the match comprises a short track and a tracked pedestrian;
(3) If the tracked pedestrian state is confirmed, turning to (4); otherwise, switching to (6);
(4) Calculating the confidence coefficient of the short track and the appearance difference of the short track and the tracked pedestrian, judging whether the short track and the tracked pedestrian meet the threshold value, and if so, turning to (5); otherwise, turning to the step (6);
(5) Repairing the features (1536-dimensional feature vectors) of the short track to obtain repaired features, and adding the repaired features into a feature pool of the tracked pedestrian; and (4) completing the updating of the feature pool, saving the feature pool, and turning to the step (2).
In the tracking module, a virtual short track algorithm based on confidence coefficient is as follows:
(1) The input of the confidence-based virtual short track algorithm is a current tracked pedestrian, the short track transmitted by the short track generation module is matched with the successfully tracked pedestrian and the short track obtained by the pedestrian tracking method based on the block chain, and the matching is called as a matching set hereinafter;
(2) If the matching set is empty, the step (7) is carried out; otherwise, taking out a match from the matching set in sequence, wherein the match comprises a short track and a tracked pedestrian;
(3) If the tracked pedestrian state is not deleted, turning to the step (4); otherwise, turning to the step (6);
(4) Calculating the confidence coefficient of the short track and the appearance difference between the short track and the tracked pedestrian, judging whether the short track and the tracked pedestrian meet the threshold value, and if so, switching to the step; otherwise, turning to the step (6);
(5) Detecting the historical track of the tracked pedestrian, judging which frames are broken, namely the tracked pedestrian has no frame in which frames, and recording the frames;
(6) Interpolating virtual frames of all recorded frames by using an interpolation algorithm according to the frames of the front and rear frames; after the interpolation is finished, the step (2) is carried out;
(7) And (5) ending the algorithm and exiting.
Compared with the prior art, the invention has the advantages that:
(1) The method designs the co-occurrence constraint network, and inhibits the shielding in the short track by utilizing the constraint of the space-time information, thereby obtaining more accurate appearance characteristics and enabling the short track to be more accurately matched, and the method has the advantages of simple network deployment and high short track extraction efficiency;
(2) The invention establishes a cross-camera pedestrian tracking method based on a block chain, which is based on a block chain technology, has less computing resource consumption, is suitable for deployment of edge computing environment, and not only has higher tracking accuracy rate, but also has faster processing speed;
(3) Aiming at the problem of long-term shielding during pedestrian tracking, the invention provides a characteristic repairing algorithm based on time simulated annealing, and the algorithm can repair the shielded area in the appearance characteristic, thereby avoiding the influence of shielding on pedestrian tracking and having the advantages of remarkably improving the tracking accuracy rate in a dense scene;
(4) Aiming at the problem of track breakage during pedestrian tracking, the invention provides a confidence coefficient-based virtual short track algorithm, and the algorithm establishes a virtual short track for the tracked pedestrian through the change of confidence coefficient and characteristics, thereby avoiding the influence of track breakage on pedestrian tracking and having the advantage of improving the integrity of target tracking.
Drawings
FIG. 1 is a flow chart of the system of the present invention;
FIG. 2 is a diagram of a co-occurrence constrained network architecture designed by the present invention;
FIG. 3 is a short track generation module of the present invention;
FIG. 4 is a tracking module of the present invention;
FIG. 5 is a flow chart of a time-based simulated annealing feature repair algorithm for the trace module of the present invention;
FIG. 6 is a flow chart of a confidence-based virtual short trajectory algorithm of the tracking module of the present invention.
Detailed Description
The present invention will be described in detail with reference to the accompanying drawings and examples.
As shown in fig. 1, the block chain-based cross-camera pedestrian tracking system of the present invention includes: the system comprises an entrance module, a data collection module, a short track generation module, a tracking module, an output module and a block chain sharing module;
the specific implementation process of each module is as follows:
1. inlet module
(1) The user inputs the name of the user name and the password through the text input box, and the system searches whether the information corresponding to the name of the user name is consistent with the input of the user in the database; if the information is consistent, the user logs in successfully and returns all information I of the user; if the information is inconsistent, prompting that the password is wrong or the user does not exist, and exiting the module;
(2) Acquiring the authority level of the user from the I, if the authority level is an administrator, opening a data export interface for the user, and displaying data export options in an operation interface; if the user is a common user, only an online viewing function is provided, and the operation interface only displays viewing options;
(3) And presenting a final operation interface, and exiting the module after the user finishes the operation.
2. The data collection module is used for shooting images of a monitored scene through the camera and acquiring continuous Video image frames Video in a monitored space within 24 hours from a lead-out interface provided by the camera; acquiring camera number c i Number the camera c i And the Video image frame Video is transmitted to a block chain module and a short track generation module.
3. The short track generation module, as shown in figure 3,
(1) The input of the short track generation module is a Video image frame Video transmitted by the data collection module, and when the number of frames contained in the Video image frame Video reaches 8, the step (2) is carried out; otherwise, continuing to wait for a new video image frame;
(2) Dividing a Video image frame Video into 8 separate frames F 1 ,…,F 8 Processing 8 independent frames by using a YOLOV3 detection network to obtain a plurality of pedestrian detection frames dets = { det = 1 ,…,det N N is the number of detection frames;
(3) Cutting 8 independent frames according to the obtained pedestrian detection frames dets, and cutting the areas patches = { patch = 1 ,…,patch N N is the number of the areas under cutting and detection sheet maps detected by packing pedestrian detection frames dets into pedestrians = { detection = i ＝(det i ,patch i )|i∈[1,N]}；
(4) Clustering the pedestrian detection slice images by using a clustering method to generate a plurality of short tracks with the minimum length of 3 and the maximum length of 8 = { T = } 1 ,…,T M T represents a short track, M represents the number of short tracks;
(5) Feature features featurees = { f > for all short traces extracted using co-occurrence constrained network 1 ,…,f M And transmitting the short traces and appearance features of the short traces into a tracking module.
The structure of the co-occurrence constrained network is shown in fig. 2, the co-occurrence constrained network receives a short track as an input, and each frame in the short track passes through the same backbone network. In fig. 2, the bone network consists of one convolutional layer, one pooling layer, and four residual layers, and the size of the output feature h is (2048 × 24 × 9). The feature h is uniformly divided into 6 blocks, each block is denoted as
Performing convolution dimensionality reduction on each block by using a 1 x 1 convolution kernel, and then obtaining a compression characteristic by using global average pooling
All the compressed features are transmitted into the Attention layer to obtain the individual features l i And spatial co-occurrence constraints S calculated from the attention mechanism i . At this time, the spatial co-occurrence is constrained by S i Transmitting the time dimension characteristic into an LSTM layer formed by stacking a plurality of LSTM units (LSTM-cells) for calculation, and designing two LSTM layers in a co-occurrence constraint networkEach followed by a random discard layer. The individual features and the output of the second immediately discarded layer are calculated in the softmax layer to obtain a set of temporal weights Tw, and a weighted average is calculated using the weights Tw and the individual features to obtain the final feature f of the short trajectory.
4. The tracking module, as shown in figure 4,
(1) The tracking module receives the short tracks tracklets = { T = delivered by the short track generation module 1 ,…,T M And appearance features of short tracks featurees = { f = } 1 ,…,f M As an input;
(2) If the number of the pedestrians h tracked currently is 0, turning to (6); otherwise, the current status of all tracked pedestrians is updated: if the current state of the tracked pedestrian is sensitive and the hit frequency hit is more than or equal to 2, changing the state of the tracked pedestrian into confirmation; if the current state of the tracked pedestrian is confirmation and the number of times of miss is greater than or equal to 18, changing the state of the tracked pedestrian into deletion; if the current state of the tracked pedestrian is sensitive, the number of times of miss is greater than or equal to 3 and the number of times of hit is equal to 0, changing the state of the pedestrian into deletion;
(3) Establishing a similarity matrix, and calculating a feature pool FP = { p) of the currently tracked pedestrian 1 ,…,p k Cosine distance between the feature and the appearance of the short track
Establishing a similarity matrix by taking the obtained cosine distance as a basic element; wherein k is the number of the features in the feature pool, p is the features in the feature pool, and the specific calculation formula and the similarity matrix are established as shown in the following formula:
in the formula, i and j represent the characteristic pool of the ith tracked pedestrian, j represents the appearance characteristic of the jth short track, H represents the total number of the current tracked pedestrians, and M represents the number of the short tracks.
(4) Filtering the similarity matrix to calculate the position P of the current tracked pedestrian i And position P 'of short track' j Is a distance of
The cosine distance between the tracked pedestrian with the distance larger than the threshold value and the short track in the similarity matrix is set to be 0; wherein, i, j in the formula represents the position of the ith tracked pedestrian, j represents the position of the jth short track, and the specific calculation formula is as follows:
(5) Distribution and optimization solution of a similarity matrix is a bipartite graph problem, and is solved by using a Hungarian algorithm, so that A = { (h) of the matching of a plurality of successfully tracked pedestrians and short tracks is obtained i ,T j ) 8230, the remaining are unsuccessfully matched short tracks and unsuccessfully matched tracked pedestrians;
(6) Initializing a new tracked pedestrian, initializing an unrefreshed pedestrian with an unsuccessfully matched short track, wherein the state of the new tracked pedestrian is sensitive, adding an appearance characteristic to the track into a characteristic pool of the new tracked pedestrian, the position of the new tracked pedestrian is an average value of frames of the short track, the miss times and the hit times of all the frames of which the historical track is the short track are initialized to 0, and the serial number ID is the maximum serial number Cur _ Max _ ID plus 1 of the currently tracked pedestrian; every time a new tracked pedestrian is initialized, adding 1 to the Cur _ Max _ ID;
(7) Updating the characteristic pool of the tracked pedestrian by using a characteristic repairing algorithm based on time simulated annealing, repairing the appearance characteristic of a short track matched with the confirmed tracked pedestrian by using the algorithm to obtain a repaired characteristic rf, and adding the repaired characteristic rf into the confirmed tracked pedestrian characteristic pool;
(8) Adding virtual short tracks dt to all tracked pedestrians by using a confidence coefficient-based virtual short track algorithm, and adding the virtual short tracks dt to historical tracks of the tracked pedestrians;
(9) Detecting whether a block chain module has a request for fusing serial numbers, and if so, modifying the serial number of the currently tracked pedestrian according to the serial number contained in the request; if not, turning to (10);
(10) If a command of finishing the tracking is received, filtering the tracked pedestrians according to the ratio of the number of the virtual short tracks to the number of the real tracks, deleting the pedestrians higher than a threshold value, storing the rest pedestrians in a database, and finishing the algorithm; if the command of ending the tracking is not received, the process goes to (1) and the next cycle is started.
5. The time-based simulated annealing feature repair algorithm of the trace module, as shown in figure 5,
(1) The input of the characteristic repairing algorithm based on the time simulated annealing is the currently tracked pedestrian h, the characteristic pool FP of the tracked pedestrian, and the short track tracklets = { T } transmitted by the short track generating module 1 ,…,T M And matching of successfully tracked pedestrians and short tracks obtained by the pedestrian tracking method based on the block chain A = { (h) i ,T j ) \8230 } (hereinafter referred to as matching set), wherein i, j are subscripts;
(2) If the matching set is empty, turning to (7); otherwise, a match is taken from the match set in order, including the short trajectory T j And tracked pedestrian h i ；
(3) If the state of the tracked pedestrian is confirmed, turning to (4); otherwise, turning to (6);
(4) Obtaining a short trajectory T j Of the appearance feature f j Calculating confidence of short track
And appearance differences of short-track and tracked pedestrians
p' is the tracked pedestrian h i Last feature in the pool of features, anWhether the two satisfy
And dif is not less than theta dif If the threshold values are met, the step (2) is carried out; otherwise, turning to (6); wherein theta is c And theta dif Are all threshold values, theta, from multiple experimental results c =0.7 and theta dif ＝0.55；
(5) Repairing the characteristics of the short track to obtain repaired characteristics rf, and adding the repaired characteristics to a characteristic pool of the tracked pedestrian, wherein the repairing formula is as follows;
where D is the number of features in the pool, p i In the representation of the ith characteristic, tau is a temperature parameter and is set to 1, and e is a natural logarithm.
(6) Completing updating of the feature pool, storing the feature pool, and turning to the step (2);
(7) And (5) finishing the algorithm and exiting the loop.
6. The confidence-based virtual short trajectory algorithm of the tracking module, as shown in figure 6,
(1) The input of the confidence coefficient-based virtual short track algorithm is the current tracked pedestrian h, the feature pool FP of the tracked pedestrian and the historical track history = { P } of the tracked pedestrian 1 \8230 } short tracks tracklets = { T = short tracks delivered by short track generation module 1 ,…,T M And matching of successfully tracked pedestrians and short tracks obtained by the pedestrian tracking method based on the block chain A = { (h) i ,T j ) \8230 } (hereinafter referred to as matching set), wherein i, j are subscripts;
(2) If the matching set is empty, turning to (7); otherwise, a match is taken from the match set in order, including the short trajectory T j And tracked pedestrian h i ；
(3) If the state of the tracked pedestrian is confirmed, turning to (4); otherwise, switching to (6);
(4) Obtaining a short trajectory T j Of the appearance feature f j Calculating confidence of short trajectory
And appearance differences of short-track and tracked pedestrians
p' is the tracked pedestrian h i The last feature in the feature pool is judged whether the last feature and the last feature meet the requirement
And dif is not less than theta dif If the threshold values are met, the step (2) is carried out; otherwise, turning to (6);
(5) Judging which frames in the feature pool do not have corresponding positions P i According to the position P of the preceding and following frames i-1 And P i+1 Interpolation is carried out to complement P i ；
(6) Completing interpolation, and turning to (2);
(7) And (5) finishing the algorithm and exiting the loop.
7. And the block chain module is responsible for sharing the data of the tracked pedestrians among different cameras. The module detects the state change of the tracking module tracking the pedestrian, if the state of one tracked pedestrian in the tracking module is changed into deletion, the characteristic pool FP and the serial number of the pedestrian are added with the serial number c of the current camera i Packing into a block and sending the block chain into a block chain; the module simultaneously detects whether a block chain has a new block n _ block continuously, when the new block n _ block is uploaded to the block chain, the data in the new block n _ block is downloaded, and if the camera number c 'stored in the new block n _ block is detected' i Different from current cameras c i Downloading a feature pool FP 'and the serial number ID' of the pedestrian in the new block, and performing pedestrian re-identification on the feature pool and all pedestrians tracked by the current tracking module by using a ReID network, if the ReID network is availableAnd if the characteristic pool is judged to be consistent with a certain tracked pedestrian, sending a fusion serial number request req to the tracking module, wherein the request comprises the serial number ID' downloaded from the new block and the serial number ID of the pedestrian tracked by the corresponding current tracking module.
Claims (5)
1. A camera-spanning pedestrian tracking system based on blockchains, comprising: the system comprises an entrance module, a data collection module, a short track generation module, a tracking module, an output module and a block chain sharing module; wherein:
the entrance module provides a management interface, so that a user can conveniently manage the system from a graphical terminal, and the user selects to enter the data collection module, the short track generation module, the tracking module, the output module and the block chain sharing module according to the actual requirement of the user; the user login interface is used, a user name and a password are input when the user logs in, and corresponding permission of the user name and the password is checked; for a common user, only online viewing permission is provided, and various data cannot be exported; for an administrator, providing all authorities, and exporting various data from a data export interface provided by the entrance module;
the data collection module is responsible for collecting data, including real-time video data of the camera, and acquiring the serial number of the current camera; shooting images of a monitored scene through a camera, and acquiring continuous video image frames in a monitored space within 24 hours from an export interface provided by the camera; transmitting the serial number of the camera to a block chain module, and transmitting the video image frame to a short track generation module;
short track generation module: firstly, receiving video image frames transmitted by a data collection module, and processing the 8 video image frames by using a YOLOV3 detection network every time 8 video image frames are received to obtain pedestrian detection frames in the 8 video image frames; according to the obtained detection frames, cutting corresponding regions from 8 video frames to form pedestrian cutting regions, and packaging the pedestrian cutting regions and the corresponding detection frames of pedestrians into detection sheet images of the pedestrians; clustering the pedestrian detection sheet image by using a clustering method to generate a short track with the minimum length of 3 frames and the maximum length of 8 frames, extracting the characteristics of the short track by using a co-occurrence constraint network after training, and finally transmitting the short track and the extracted characteristics of the short track into a tracking module;
the system comprises a tracking module, a block chain-based cross-camera pedestrian tracking method, a plurality of pedestrian tracking module and a database, wherein the plurality of data are maintained for each pedestrian, and part of the data are stored in the database, the plurality of data comprise a feature pool for storing the appearance features of the pedestrian, the number of times of miss, a historical track, the current state of the pedestrian, the serial number of the pedestrian, the number of times of hit and the current position of the pedestrian, the number of times of hit refers to the number of times that the pedestrian successfully obtains short track matching, and the number of times of miss refers to the number of frames that the pedestrian successfully obtains matching distance from the current frame last time; the current state comprises one of sensitivity, confirmation and deletion, wherein the sensitivity means that the pedestrian is possibly caused by false detection, the confirmation means that the pedestrian is being tracked, and the deletion means that the pedestrian is lost from the detection area; when the module is implemented, firstly, the short track generated by the short track generation module and the characteristics of the short track are received, then the state of the currently tracked pedestrian is updated, and a similarity matrix between the currently tracked pedestrian and the received short track is calculated; then according to the position of the current tracked pedestrian and the input short track, filtering the similarity matrix to obtain a filtered similarity matrix; the optimal distribution scheme of the filtered similarity matrix is solved by using a Hungarian algorithm, a plurality of paired matrix subscripts are obtained, namely the successfully matched pedestrians and short tracks are obtained, and the unsuccessfully matched pedestrians and short tracks which are not obtained are called unsuccessfully matched pedestrians; adding one to the number of times of miss of the unsuccessfully matched pedestrians, and initializing unsuccessfully matched short tracks as new pedestrians; after all unsuccessfully matched pedestrians and short tracks are processed, updating a feature pool of the pedestrian appearance features of which the current states are confirmed and tracked by using a feature restoration algorithm based on time simulated annealing; after the feature pools of the appearance features of all tracked pedestrians are updated, virtual short tracks are added to all tracked pedestrians by using a confidence-based virtual short track algorithm, namely the short tracks are not directly generated by a short track generation module, but are estimated according to historical tracks of the pedestrians; adding the virtual short trajectory to a historical trajectory of the tracked pedestrian; if the entrance module finishes the tracking command, storing the historical tracks of all tracked pedestrians in a database;
the block chain module is responsible for sharing the data of the pedestrians tracked among different cameras; the module detects the state change of the tracking module for tracking the pedestrian, if the state of one tracked pedestrian in the tracking module is changed into deletion, the characteristic pool and the serial number of the pedestrian as well as the serial number of the current camera are packed into a block, and the block is sent to a block chain; the module continuously detects whether a new block exists in the block chain or not, when the new block is uploaded to the block chain, data in the new block are downloaded, if the camera number stored in the new block is different from that of a current camera, serial numbers of a feature pool and pedestrians in the new block are downloaded, a ReID network is used for re-identifying all pedestrians tracked by the feature pool and the current tracking module, if the ReID network judges that the feature pool is consistent with a certain tracked pedestrian, a fused serial number request is sent to the tracking module, and the request comprises the serial number downloaded from the new block and the corresponding serial number of the pedestrian tracked by the current tracking module;
the output module provides two data interfaces, one data interface only provides the view of the historical track of the pedestrian in the database, and the other data interface not only provides the view, but also provides an interface for deriving the historical track of the pedestrian; the module provides an interface for a user to check information, and encapsulates other modules, so that each module is transparent to the user;
in the tracking module, a cross-camera pedestrian tracking method based on a block chain is as follows:
(1) Receiving the short track and the feature of the short track transmitted by the short track generation module, wherein the feature is a feature vector with 1536 dimensions;
(2) If the currently tracked pedestrian is empty, turning to the step (6); otherwise, the current status of all tracked pedestrians is updated: if the current state of the tracked pedestrian is sensitive and the hit frequency is more than or equal to 2, changing the state of the tracked pedestrian into confirmation; if the current state of the tracked pedestrian is confirmation and the number of times of miss is 18 or more, changing the state thereof to deletion; if the current state of the tracked pedestrian is sensitive, the number of times of miss is greater than or equal to 3, and the number of times of hit is equal to 0, changing the state of the tracked pedestrian into deletion;
(3) Establishing a similarity matrix, calculating the cosine distance between the features in the feature pool of the currently tracked pedestrian and the features of the short track, and establishing the similarity matrix by taking the obtained cosine distance as a basic element;
(4) Filtering a similarity matrix, calculating the distance between the position of the current tracked pedestrian and the position of the short track, and setting the cosine distance between the tracked pedestrian with the distance greater than the threshold value and the short track in the similarity matrix as 0;
(5) Distributing, wherein the optimal solution of the similarity matrix is a bipartite graph problem, the solution is carried out by using a Hungarian algorithm, the matching of a plurality of successfully tracked pedestrians and short tracks is obtained, and the rest are unsuccessfully matched short tracks and unsuccessfully matched tracked pedestrians;
(6) Initializing a new tracked pedestrian, initializing an unrefreshed pedestrian with an unsuccessfully matched short track, wherein the state of the new tracked pedestrian is sensitive, adding an appearance characteristic to the track into a characteristic pool of the new tracked pedestrian, the position of the new tracked pedestrian is an average value of frames of the short track, the miss times and the hit times of all the frames of the historical track which is the short track are initialized to 0, and the serial number is the maximum serial number of the currently tracked pedestrian plus 1;
(7) Updating the feature pool of the tracked pedestrian by using a feature repairing algorithm based on time simulated annealing, repairing the appearance feature of the short track matched with the confirmed tracked pedestrian by using the algorithm to obtain a repaired feature, and adding the repaired feature into the confirmed tracked pedestrian feature pool;
(8) Adding virtual short tracks to all tracked pedestrians by using a confidence coefficient-based virtual short track algorithm, and adding the virtual short tracks to historical tracks of the tracked pedestrians;
(9) Detecting whether a block chain module has a request for fusing serial numbers, and if so, modifying the serial number of the currently tracked pedestrian according to the serial number contained in the request; if no, turning to the step (10);
(10) If a command of ending the tracking is received, filtering the tracked pedestrians according to the ratio of the number of the virtual short tracks to the number of the real tracks, deleting the pedestrians higher than a threshold value, storing the remaining pedestrians in a database, and ending; and (4) if the command of ending the tracking sent by the entrance module is not received, the step (1) is carried out.
2. The blockchain-based cross-camera pedestrian tracking system of claim 1, wherein: in the short track generation module, the structure and training process of the co-occurrence constraint network are as follows:
(1) The co-occurrence constraint network receives a short track with the length of 8 as input, and if the given short track is less than 8, frames are inserted by using a linear interpolation method, so that the length of the short track is 8;
(2) Establishing a co-occurrence constraint network, using a network structure before a ResNet50 global average pooling layer as a backbone network, and uniformly dividing a feature graph output by the backbone network by using a fixed dividing layer at the back, wherein the feature graph is divided into 6 blocks, and each block is called a co-occurrence local feature; processing all co-occurring local features by using a 1 × 1 convolution kernel and an average pooling layer, and reducing the dimensionality of the co-occurring local features to obtain compressed features; the compressed features are input into a multi-head attention layer to calculate spatial co-occurrence constraints among the compressed features; meanwhile, a plurality of compression features are spliced into individual features; the structure from the backbone network to the completion of the splicing of the compression features is shared by all frames in the short track, that is, each frame in the short track is subjected to the same structural processing, and finally, respective individual features and respective spatial co-occurrence constraints are obtained; the spatial co-occurrence constraints of all frames are packed into a sequence and input into an LSTM layer of 256 hidden units, after the output of the LSTM layer passes through a random discarding layer, the output of the random discarding layer is input into an LSTM layer of another 256 hidden units, the output of the LSTM layer is input into another random discarding layer, the output of the random discarding layer is input into a full connection layer, the full connection output is converted into a group of weights by a sofmax layer, and the individual features of all frames are combined into short track features by using the reorganization time weights; the short track characteristics are input into a full connection layer to carry out short track classification, namely, the last full connection layer of the network outputs a label corresponding to the short track;
(3) The co-occurrence constraint network completes training on an iLIDS-VID (public data set on the Internet), video data in the data set is segmented into short tracks with the length of 8, and the short tracks generated by the same video have the same label; all the short tracks are input into a co-occurrence constraint network, and the short tracks are classified to obtain predicted labels; calculating the loss of the tags predicted by the co-occurrence constrained network and the real tags in the data set, and updating the weight parameters of all layers behind the backbone network of the co-occurrence constrained network by using a gradient descent method; after loss convergence, fixing and storing the weight parameters, and simultaneously storing the network structure of the co-occurrence constraint network;
(4) And the short track generation module loads weight parameters of the co-occurrence constraint network after pre-training, and performs feature extraction on the short tracks generated by clustering by using the network to obtain features of the short tracks.
3. The blockchain-based cross-camera pedestrian tracking system of claim 1, wherein: in the tracking module, a characteristic repairing algorithm based on time simulated annealing is realized as follows:
(1) The input of the characteristic repairing algorithm based on time simulated annealing is the current tracked pedestrian, the short track transmitted by the short track generating module, and the matching between the successfully tracked pedestrian and the short track obtained by the pedestrian tracking method based on the block chain, which is called as a matching set hereinafter;
(2) If the matching set is empty, the step (7) is carried out; otherwise, taking out a match from the matching set in sequence, wherein the match comprises a short track and a tracked pedestrian;
(3) If the tracked pedestrian state is confirmed, turning to the step (4); otherwise, turning to the step (6);
(4) Calculating the confidence coefficient of the short track and the appearance difference of the short track and the tracked pedestrian, judging whether the short track and the tracked pedestrian meet a threshold value, and if so, turning to the step (5); otherwise, turning to the step (6);
(5) Repairing the characteristics of the short tracks to obtain repaired characteristics, and adding the repaired characteristics into a characteristic pool of the tracked pedestrians; completing the updating of the feature pool, saving the feature pool, and turning to the step (2);
(6) And ending and exiting.
4. The blockchain-based cross-camera pedestrian tracking system of claim 1, wherein: in the tracking module, a virtual short track algorithm based on confidence coefficient is as follows:
(1) The input of the virtual short track algorithm based on the confidence coefficient is the current tracked pedestrian, the short track transmitted by the short track generation module and the matching of the successfully tracked pedestrian and the short track obtained by the pedestrian tracking method based on the block chain are called as a matching set in the following;
(2) If the matching set is empty, the step (7) is carried out; otherwise, taking out a match from the matching set in sequence, wherein the match comprises a short track and a tracked pedestrian;
(3) If the tracked pedestrian state is not deleted, turning to the step (4); otherwise, turning to the step (6);
(4) Calculating the confidence of the short track and the appearance difference between the short track and the tracked pedestrian, judging whether the short track and the tracked pedestrian meet a threshold value, and if both the short track and the tracked pedestrian meet the threshold value, switching to the next step; otherwise, turning to the step (6);
(5) Detecting the historical track of the tracked pedestrian, judging which frames are broken, namely the tracked pedestrian has no frame in which frames, and recording the frames;
(6) Interpolating virtual frames of all recorded frames by using an interpolation algorithm according to the frames of the front and rear frames; after the interpolation is finished, the step (2) is carried out;
(7) And ending and exiting.
5. A cross-camera pedestrian tracking method based on a block chain is characterized by comprising the following implementation steps:
(1) The user opens an entrance interface, inputs own user name and password according to the interface prompt, the entrance module checks the user name and the password after logging in, and the user is allowed to log in the system after the check is passed; opening different authorities according to authority information corresponding to a user account, if the user is an administrator, providing all authorities, opening a data export interface, and if the user is an ordinary user, only providing an online viewing function;
(2) When a user normally logs in the system, a pedestrian tracking task is started to be executed, the data collection module obtains real-time video data input and transmits the real-time video data input to the short track generation module, the short track generation module obtains detection from the obtained video data and generates a short track, meanwhile, a co-occurrence constraint network is used for completing feature extraction of the short track, and the short track and features of the short track are packaged and transmitted to the tracking module;
(3) The tracking module acquires the short track and short track characteristics transmitted by the short track generation module, tracks a plurality of pedestrians, updates the state of the currently tracked pedestrian, and calculates a similarity matrix between the currently tracked pedestrian and the received short track; then according to the position of the current tracked pedestrian and the input short track, filtering the similarity matrix to obtain a filtered similarity matrix; the optimal distribution scheme of the filtered similarity matrix is solved by using a Hungarian algorithm, and the successfully matched pedestrians and short tracks are obtained; adding one to the number of times of miss of the unsuccessfully matched pedestrians, and initializing unsuccessfully matched short tracks as new pedestrians; after all unsuccessfully matched pedestrians and short tracks are processed, updating a feature pool of which the current states are confirmed and tracked pedestrian appearance features by using a feature repairing algorithm based on time simulated annealing; after the feature pools of the appearance features of all tracked pedestrians are updated, virtual short tracks are added to all tracked pedestrians by using a confidence-based virtual short track algorithm, namely the virtual short tracks are added to the historical tracks of the tracked pedestrians according to the short tracks estimated by the historical tracks of the pedestrians instead of being generated by a short track generation module; if the entrance module finishes the tracking command, storing the historical tracks of all tracked pedestrians in a database;
(4) The block chain module and the tracking module start to start in the system at the same time to detect the state change of the tracked pedestrian, and if the state of a certain tracked pedestrian in the tracking module is deleted, the information of the pedestrian is packaged and sent to the block chain; the module continuously detects whether a new block exists in the block chain, when the new block is uploaded to the block chain, data in the new block are downloaded, if the camera number stored in the new block is different from that of a current camera, serial numbers of a feature pool and pedestrians in the new block are downloaded, a user re-identification ReID network is used for re-identifying all pedestrians tracked by the feature pool and the current tracking module, and if the ReID network judges that the feature pool is consistent with a certain tracked pedestrian, a fusion serial number request is sent to the tracking module, wherein the request comprises the serial number downloaded from the new block and the serial number of the corresponding pedestrian tracked by the current tracking module;
(5) And the user inlet module sends a tracking finishing instruction, and the output module opens a corresponding interface according to the user authority for the user to export data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010783800.0A CN112131929B (en) | 2020-08-06 | 2020-08-06 | Cross-camera pedestrian tracking system and method based on block chain |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010783800.0A CN112131929B (en) | 2020-08-06 | 2020-08-06 | Cross-camera pedestrian tracking system and method based on block chain |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112131929A CN112131929A (en) | 2020-12-25 |
| CN112131929B true CN112131929B (en) | 2022-10-11 |
Family
ID=73850813
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010783800.0A Active CN112131929B (en) | 2020-08-06 | 2020-08-06 | Cross-camera pedestrian tracking system and method based on block chain |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112131929B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112668487B (en) * | 2020-12-29 | 2022-05-27 | 杭州晨安科技股份有限公司 | Teacher tracking method based on fusion of body fitness and human similarity |
| CN112861785B (en) * | 2021-03-09 | 2023-09-29 | 中山大学 | Instance segmentation and image restoration-based pedestrian re-identification method with shielding function |
| CN113206980B (en) * | 2021-04-08 | 2022-06-07 | 北京航空航天大学 | Intelligent node recommendation system in end edge cloud fusion crowd sensing environment |
| CN116542858B (en) * | 2023-07-03 | 2023-09-05 | 众芯汉创(江苏)科技有限公司 | Data splicing analysis system based on space track |
| CN117435934B (en) * | 2023-12-22 | 2024-07-19 | 中国科学院自动化研究所 | Matching method, device and storage medium of moving target track based on bipartite graph |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105069408B (en) * | 2015-07-24 | 2018-08-03 | 上海依图网络科技有限公司 | Video portrait tracking based on recognition of face under a kind of complex scene |
| CN108875588B (en) * | 2018-05-25 | 2022-04-15 | 武汉大学 | Cross-camera pedestrian detection tracking method based on deep learning |
| CN110427904B (en) * | 2019-08-08 | 2022-03-04 | 南京理工大学 | Mall monitoring system, method and device based on pedestrian re-identification |
-
2020
- 2020-08-06 CN CN202010783800.0A patent/CN112131929B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN112131929A (en) | 2020-12-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112131929B (en) | Cross-camera pedestrian tracking system and method based on block chain | |
| Chen et al. | An edge traffic flow detection scheme based on deep learning in an intelligent transportation system | |
| CN112001339B (en) | Pedestrian social distance real-time monitoring method based on YOLO v4 | |
| CN107330396B (en) | Pedestrian re-identification method based on multi-attribute and multi-strategy fusion learning | |
| CN105844239B (en) | It is a kind of that video detecting method is feared based on CNN and LSTM cruelly | |
| CN111161315B (en) | Multi-target tracking method and system based on graph neural network | |
| CN109784269A (en) | One kind is based on the united human action detection of space-time and localization method | |
| CN110222592B (en) | Construction method of time sequence behavior detection network model based on complementary time sequence behavior proposal generation | |
| CN110084165A (en) | The intelligent recognition and method for early warning of anomalous event under the open scene of power domain based on edge calculations | |
| CN109033107A (en) | Image search method and device, computer equipment and storage medium | |
| CN110457984A (en) | Pedestrian's attribute recognition approach under monitoring scene based on ResNet-50 | |
| CN110163069A (en) | Method for detecting lane lines for assisting driving | |
| CN114821374B (en) | Knowledge and data collaborative driving unmanned aerial vehicle aerial photography target detection method | |
| CN116343330A (en) | Abnormal behavior identification method for infrared-visible light image fusion | |
| EP4105828A1 (en) | Model updating method and related device | |
| CN114550053A (en) | Traffic accident responsibility determination method, device, computer equipment and storage medium | |
| CN107480618A (en) | A kind of data analysing method of big data platform | |
| CN113870312B (en) | Single target tracking method based on twin network | |
| CN111126155A (en) | Pedestrian re-identification method for generating confrontation network based on semantic constraint | |
| CN115205786A (en) | On-line automatic identification and alarm method for mobile phone pirate behavior | |
| CN115187906A (en) | Pedestrian detection and re-identification method, device and system | |
| CN115909192A (en) | Pedestrian detection method based on improved EfficientDet | |
| Nan et al. | Safety Helmet Detection Dynamic Model Based on the Critical Area Attention Mechanism | |
| CN114581769A (en) | Method for identifying houses under construction based on unsupervised clustering | |
| CN112699846B (en) | Specific character and specific behavior combined retrieval method and device with identity consistency check function |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |