A target tracking method and system based on face clustering

Abstract

The present invention discloses a target tracking method and system based on face clustering, which belongs to the field of face recognition technology. First, a personnel database is created and personnel information is input; then, features are extracted through the face pictures captured by the front-end face capture machine; the distance measurement method of the class center feature value is adopted to calculate the difference between each face feature value, and the face pictures captured by the front-end face capture machine are compared with the face pictures in the personnel database and processed accordingly, and then feature values ​​are extracted according to the face pictures uploaded by the system and compared with the cover pictures of the personnel files, and the personnel passage records whose similarity reaches the set threshold are retrieved, and the target motion trajectory is dynamically drawn. Aiming at the complex structure, dynamic characteristics, massive data, and diversification of large-scale face data at the city level, the present invention improves the parallelism of MapReduce task processing by designing a differential storage strategy of clustering, realizes rapid data retrieval, and improves the efficiency of data processing and the real-time performance of result display.

Description

Technical Field

[0001] The present invention relates to the technical field of face recognition, and in particular to a target tracking method and system based on face clustering. Background Art

[0002] In recent years, with the high degree of informatization in all walks of life, data is growing at an unprecedented rate. The era of big data has arrived. The massiveness and complexity of big data require that the indexing mechanism of big data must meet the requirements of supporting multiple queries, supporting efficient retrieval and easy maintenance.

[0003] With the rapid development of video surveillance hardware and software and face capture technology, face data is increasing rapidly, which also brings challenges to its application. Face recognition is to extract features for recognition from face images according to a certain strategy, that is, to map data in the face space to the feature space. The recognition of large-scale face databases has become difficult to be practical due to the long comparison time caused by the linear growth of face data, which makes it difficult to efficiently complete the confidence and serialization of portrait files and the rapid retrieval of personnel trajectories in large-scale portrait databases. How to perform cluster analysis on big face data and extract valuable knowledge is a problem that needs to be solved urgently in current research. To this end, a target tracking method and system based on face clustering is proposed. Summary of the invention

[0004] The technical problem to be solved by the present invention is: how to address the problem of low retrieval efficiency of large-scale face libraries, and provide a target tracking method based on face clustering. This method can perform cluster analysis on faces and realize a rapid retrieval method for personnel trajectories in large-scale portrait libraries.

[0005] The present invention solves the above technical problems through the following technical solutions, and the present invention comprises the following steps:

[0006] S1: Enter facial images and personnel-related information to establish a personnel database, i.e., the base database;

[0007] S2: Extract feature values ​​from the input face image;

[0008] S3: Establish a facial image capture database and extract feature values ​​from facial images captured by a facial capture machine;

[0009] S4: Use the distance measurement method of the class center feature value to calculate the difference between the face feature values, and construct an index tree by designing a clustering difference storage strategy;

[0010] S5: For the face images whose face feature value similarity reaches the threshold, the face images are archived and the ID number is used as the ID;

[0011] S6: For the face images whose facial feature value similarity does not reach the threshold, the system automatically creates an ID for the face images;

[0012] S7: Establish a personnel movement trajectory analysis model to analyze the personnel movement records in the selected time area;

[0013] S8: Retrieve results by comparing and analyzing the classified personnel files;

[0014] S9: Dynamically draw the movement trajectory of the retrieved person snapshot records in chronological order based on the map.

[0015] Furthermore, in step S1, the person-related information includes facial image, ID number, name, gender, and address.

[0016] Furthermore, in steps S2 and S3, feature value extraction refers to constructing a unified feature vector representation rule by analyzing the face image description corresponding to each feature in the face picture geometric feature vector provided by the face capture machine, and realizing feature value extraction of face data through a convolutional neural network.

[0017] Furthermore, in step S4, the distance measurement method of the class center feature value comprehensively considers the distinguishability and independence of the facial feature vector, adopts the distance measurement scheme of the class center feature value, calculates the difference between each facial feature value, and compares the extracted facial features with the existing class center features. First, search through short features to filter out dissimilar classes and narrow the search range, and then search through long features to finally determine the class to which the facial features belong. After the facial features are included in the class, the class center features are recalculated, and one class represents a person.

[0018] Furthermore, in step S4, clustering is implemented based on similarity. In each cluster, the clusters are layered into concentric circles with the center of mass as the center, based on the idea that the closer the point is to the centroid, the greater the probability of being queried. Each layer is expressed using binary codes of different lengths, and an index tree is constructed using the codes to achieve fast data retrieval.

[0019] Furthermore, in step S7, the motion trajectory analysis model refers to establishing a personnel motion trajectory analysis model using MapReduce under the Hadoop platform. After the clustering task is completed, the classes and their class features are extracted, and the class features are used as input to perform a 1:N comparison on the personnel database, and obtain the first base database personnel whose similarity is greater than the similarity threshold, associate this class with the identity ID of the base database personnel, retrieve the personnel snapshot records, and dynamically draw the trajectory based on Canvas technology to display it to the user.

[0020] The present invention also discloses a target tracking system based on face clustering, which uses the above tracking method to implement target tracking based on face clustering, including:

[0021] Personnel database management module: used to import a single face image from outside the system, or batch import multiple face images, and supports editing of face image information;

[0022] Snapshot library management module: used to extract feature values ​​from imported images and support retrieval operations based on time and area conditions;

[0023] Face clustering module: It is used to dynamically cluster face feature values ​​and establish personnel files by using the distance measurement method of class center feature values;

[0024] Personnel movement trajectory analysis module: used to retrieve personnel snapshot records according to the selected time, area and uploaded face pictures, and draw movement trajectories based on maps;

[0025] Control processing module: used to issue instructions to other modules to complete related actions;

[0026] The personnel database management module, the snapshot database management module, the face clustering module, and the personnel movement trajectory analysis module are all connected to the control processing module.

[0027] Compared with the prior art, the present invention has the following advantages: a target tracking method based on face clustering, targeting the complex structure, dynamic characteristics, massive data, diversity, and inability to sort of large-scale city-level face data, by designing a differential storage strategy for clustering, improving the parallelism of MapReduce task processing, and achieving rapid data retrieval, greatly improving the efficiency of data processing and the real-time display of results, and is worthy of promotion and use. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a flow chart of a target tracking method in an embodiment of the present invention;

[0029] Figure 2 is a schematic diagram of the process of clustering facial feature values ​​in an embodiment of the present invention;

[0030] Figure 3 It is a flow chart of the personnel movement trajectory analysis model established in an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiment.

[0032] This embodiment provides a technical solution: a target tracking method based on face clustering, including the following steps: first, create a personnel database and enter personnel-related information, such as face pictures, ID card numbers, names, genders, addresses, etc.; second, extract the feature values ​​of face data by analyzing the face image description corresponding to each feature in the face picture geometric feature vector provided by the face capture machine; then, use the distance measurement scheme of the class center feature value to calculate the difference between each face feature value, compare the face picture captured by the front-end face capture machine with the face picture in the personnel database, and when the similarity reaches the set threshold, establish the personnel file with the ID number as the ID and classify the captured picture into the file. If the set threshold is not reached, the system automatically generates an ID and establishes the personnel file. Finally, extract the feature value based on the face picture uploaded by the system and compare it with the cover picture of the personnel file, retrieve the personnel passage records whose similarity reaches the set threshold, and dynamically draw the target motion trajectory based on GIS.

[0033] like Figure 1 As shown, the present invention adopts a target tracking method based on face clustering, comprising:

[0034] 1) Input facial images and personnel-related information to establish a personnel database;

[0035] 2) Extract feature values ​​from the input face image;

[0036] 3) Establish a facial image capture database. Extract feature values ​​from facial images captured by the facial capture machine;

[0037] 4) Face clustering. The distance measurement scheme of the class center feature value is used to calculate the difference between the face feature values. The index tree is constructed by designing a clustering difference storage strategy to achieve fast data retrieval;

[0038] 5) For face images whose facial feature value similarity reaches the threshold, they are archived and their ID number is used as the ID;

[0039] 6) For face images whose facial feature value similarity does not reach the threshold, the system will automatically create an ID for the face image;

[0040] 7) Establish a personnel movement trajectory analysis model to analyze the personnel movement records in the selected time area;

[0041] 8) Retrieve results by comparing and analyzing the classified personnel files;

[0042] 9) Dynamically draw the retrieved personnel snapshot records in chronological order based on the map.

[0043] In this embodiment, feature value extraction refers to constructing a unified feature vector representation rule by analyzing the face image description corresponding to each feature in the face picture geometric feature vector provided by the face capture machine, and realizing 128-dimensional feature value extraction of face data based on a convolutional neural network.

[0044] In this embodiment, the distance measurement technology of the class center feature value refers to comprehensively considering the distinguishability and independence of the facial feature vector, and specifically adopting the distance measurement scheme of the class center feature value to calculate the difference between each facial feature value, that is, to compare the extracted facial features with the existing class center features, first perform a rough search through short features to filter out a large number of dissimilar classes, narrow the search range, and then perform a fine search through long features to finally determine the class to which the feature belongs. After the feature is included in the class, the class center feature is recalculated to more accurately describe the class information, which can effectively avoid the interference of the descriptions between the features on the clustering results, and realize efficient dynamic clustering of facial feature values.

[0045] In this embodiment, the method for rapid data retrieval refers to designing a differential storage strategy for clustering, that is, automatic clustering based on similarity, one cluster represents a person, and clustering of massive data can greatly reduce the number of retrievals and improve the parallel processing of MapReduce tasks. In each cluster, based on the idea that the closer the point to the centroid is, the greater the probability of being queried, the clusters are layered into concentric circles with the centroid as the center, and each layer is expressed using binary codes of different lengths. The codes are used to construct an index tree to achieve rapid data retrieval.

[0046] In this embodiment, the personnel movement trajectory analysis model refers to the establishment of a personnel movement trajectory analysis model using MapReduce on the Hadoop platform. After the clustering task is completed, the classes and their class features are extracted, and the class features are used as input to perform a 1:N comparison on the personnel database, and the TOP1 personnel in the database with a similarity threshold greater than the similarity threshold are obtained. This class is associated with the unique identifier (i.e., identity ID) of the personnel in the database, and the personnel snapshot records are retrieved and the trajectory is dynamically drawn based on Canvas technology to display to the user.

[0047] like Figure 2 As shown, the distance measurement technology of the class center eigenvalue is used to calculate the difference between each face eigenvalue, so as to realize efficient dynamic clustering of face eigenvalues. The specific steps include:

[0048] 1) Extract the feature values ​​of the face base database;

[0049] 2) Extract feature values ​​from the face image provided by the connected face capture machine;

[0050] 3) Using the distance measurement method of the class center feature value to calculate the feature value difference between the face image provided by the face capture machine and the face base image;

[0051] 4) Archive the face image based on the comparison result. If the set threshold is reached, a personnel file with the person's ID card as the ID is created. If the set threshold is not reached, a file of the person's ID is automatically generated.

[0052] like Figure 3 As shown, the present invention also provides a target tracking system based on face clustering, which is used to run the above-mentioned target tracking method based on face clustering, including:

[0053] (1) Personnel database management module: It can import a single face image from outside the system, or batch import multiple face images, and supports operations such as editing face image information;

[0054] (2) Snapshot library management module: extracts feature values ​​from imported images and supports retrieval based on time, region, and other conditions;

[0055] (3) Face clustering module: It uses the distance measurement scheme of the class center feature value to achieve efficient dynamic clustering of face feature values ​​and establish personnel files;

[0056] (4) Personnel movement trajectory analysis module: retrieves personnel snapshot records based on the selected time, area and uploaded face images, and draws movement trajectories based on the map.

[0057] In summary, the target tracking method based on face clustering in the above embodiment aims at the complex structure, dynamic characteristics, massive data, diversity, and inability to sort of large-scale face data at the city level. By designing a differential storage strategy for clustering, the parallelism of MapReduce task processing is improved, and fast data retrieval is achieved. This greatly improves the efficiency of data processing and the real-time display of results, and is worthy of promotion and use.

[0058] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A target tracking method based on face clustering, characterized in that: The following steps are involved: S1: Enter facial images and personnel-related information to establish a personnel database, i.e., the base database; S2: Extract feature values ​​from the input face image; S3: Establish a facial image capture database and extract feature values ​​from facial images captured by a facial capture machine; S4: Use the distance measurement method of the class center feature value to calculate the difference between the face feature values, and construct an index tree by designing a clustering difference storage strategy; In the step S4, the distance measurement method of the class center feature value is to comprehensively consider the distinguishability and independence of the face feature vector, adopt the distance measurement scheme of the class center feature value, calculate the difference between each face feature value, that is, compare the extracted face feature with the existing class center feature, first search through the short feature, filter out the dissimilar class, narrow the search range, and then search through the long feature to finally determine the class to which the face feature belongs, and recalculate the class center feature after the face feature is included in the class, where one class represents a person; S5: For the face images whose face feature value similarity reaches the threshold, the face images are archived and the ID number is used as the ID; S6: For the face images whose facial feature value similarity does not reach the threshold, the system automatically creates an ID for the face images; S7: Establish a personnel movement trajectory analysis model to analyze the personnel movement records in the selected time area; S8: Retrieve results by comparing and analyzing the classified personnel files; S9: Dynamically draw the movement trajectory of the retrieved person snapshot records in chronological order based on the map.

2. The target tracking method based on face clustering according to claim 1, characterized in that: In step S1, the person-related information includes facial image, ID number, name, gender, and address.

3. The target tracking method based on face clustering according to claim 2, characterized in that: In steps S2 and S3, feature value extraction refers to constructing a unified feature vector representation rule by analyzing the face image description corresponding to each feature in the face picture geometric feature vector provided by the face capture machine, and realizing feature value extraction of face data through a convolutional neural network.

4. The target tracking method based on face clustering according to claim 3, characterized in that: In step S4, clustering is implemented based on similarity. In each cluster, the clusters are layered into concentric circles with the center of mass as the center, based on the idea that the closer the point is to the centroid, the greater the probability of being queried. Each layer is expressed using binary codes of different lengths, and an index tree is constructed using the codes to implement data retrieval.

5. The target tracking method based on face clustering according to claim 4, characterized in that: In step S7, the motion trajectory analysis model refers to establishing a personnel motion trajectory analysis model using MapReduce under the Hadoop platform. After the clustering task is completed, the classes and their class features are extracted, and the class features are used as input to perform a 1:N comparison on the personnel database, and obtain the first base database personnel with a similarity greater than the similarity threshold, associate this class with the identity ID of the base database personnel, retrieve the personnel snapshot records, and dynamically draw the trajectory based on Canvas technology to display it to the user.

6. A target tracking system based on face clustering, characterized in that: The tracking method according to any one of claims 1 to 5 is used to implement target tracking based on face clustering, comprising: Personnel database management module: used to import a single face image from outside the system, or batch import multiple face images, and supports editing of face image information; Snapshot library management module: used to extract feature values ​​from imported images and support retrieval operations based on time and area conditions; Face clustering module: It is used to dynamically cluster face feature values ​​and establish personnel files by using the distance measurement method of class center feature values; Personnel movement trajectory analysis module: used to retrieve personnel snapshot records according to the selected time, area and uploaded face pictures, and draw movement trajectories based on maps; Control processing module: used to issue instructions to other modules to complete related actions; The personnel database management module, the snapshot database management module, the face clustering module, and the personnel movement trajectory analysis module are all connected to the control processing module.

Images