A scene image segmentation system based on local feature fusion

Abstract

The present invention discloses a scene image segmentation system based on local feature fusion, relates to the field of image segmentation technology, and effectively improves the extensiveness of scene object recognition. The present invention constructs each scene target in an application scene according to a scene data set, extracts multi-source data of each scene target from the scene data set, and then classifies the scene targets into static scene targets and dynamic scene targets, constructs a static scene image model and several multi-source dynamic scene image models of the application scene according to the multi-source data of the static scene targets and the dynamic scene targets, and sets multiple main recognition points and secondary recognition points on the multi-source dynamic scene image model, generates several real-time scene target image models according to the real-time scene data set, matches the real-time scene target image model with the multi-source dynamic scene image model and the static scene image model, and recognizes each real-time scene target image model according to the matching result.

Description

Technical Field

[0001] The invention belongs to the technical field of image segmentation, and in particular relates to a scene image segmentation system based on local feature fusion. Background Art

[0002] Scene object recognition is a computer vision technology that aims to identify and classify different objects in images or videos. This technology can be applied in a variety of scenarios, including intelligent surveillance, self-driving cars, augmented reality applications, smart homes, and industrial automation.

[0003] Existing scene object recognition technology has the following defects:

[0004] Insufficient accuracy: Although object recognition technology has made great progress, it still has the problem of insufficient accuracy in complex scenes. In particular, when the object is partially blocked, the lighting conditions are poor, or there are similar objects, the recognition accuracy may decrease.

[0005] Limited adaptability to diverse scenarios: Object recognition techniques are usually trained in specific scenarios and therefore may lack sufficient adaptability to new and diverse scenarios. This may lead to poor recognition performance in some novel scenarios.

[0006] Therefore, how to improve the extensiveness of scene object recognition while ensuring recognition accuracy is a difficulty in existing technologies. Summary of the invention

[0007] In order to solve the above technical problems, the present invention provides a scene image segmentation system based on local feature fusion, including a cloud computing platform, wherein the cloud computing platform is communicatively connected with a scene data acquisition module, a scene target segmentation module, a scene target analysis module and a target recognition module;

[0008] The scene data acquisition module is used to set several types of sensors in the application scene and set several data acquisition cycles, and then call the sensors to collect scene data in each data acquisition cycle, and generate corresponding scene data sets to send to the scene target segmentation module;

[0009] The scene target segmentation module is used to construct each scene target in the application scene according to the scene data set, and extract multi-source data of each scene target from the scene data set, and then classify the scene targets into static scene targets and dynamic scene targets;

[0010] The scene target analysis module is used to construct a static scene image model and several multi-source dynamic scene image models of the application scene according to the multi-source data of the static scene target and the dynamic scene target, and set a plurality of primary identification points and secondary identification points on the multi-source dynamic scene image model;

[0011] The target recognition module is used to generate a plurality of real-time scene target image models according to the real-time scene data set, match the real-time scene target image models with multi-source dynamic scene image models and static scene image models, and recognize each real-time scene target image model according to the matching results.

[0012] Furthermore, the process of collecting the scene data set includes:

[0013] The scene data acquisition module divides the target scene into N scene sub-areas of the same size, and sets a number for each scene sub-area, where N is a positive integer greater than 0;

[0014] A plurality of sensors of the same type are arranged in each scene sub-area, and the same data collection cycle is set for all sensors. When the data collection cycle starts, the scene data collection module sends a data collection instruction to each sensor at the same time, and then each sensor collects various scene data of the scene sub-area in which it is located according to the data collection instruction, wherein the scene data includes scene optical video data, scene infrared video data, etc.;

[0015] When the data collection cycle is over, each sensor uploads the collected data to the scene data collection module, and then the scene data collection module marks the corresponding scene sub-area number for each scene data, and integrates the scene data with the same number to generate a scene data set of the corresponding scene sub-area, and marks the corresponding scene sub-area number;

[0016] At the same time, community or factory managers upload residents' or workers' identity information to the scene data collection module. The resident or worker identity information includes the name, identity number and photo of the people entering and leaving the corresponding application scene. The scene data collection module then generates a scene personnel identity data set based on the resident or worker identity information.

[0017] Furthermore, the process of constructing the scene target includes:

[0018] The scene object segmentation module obtains the scene data of each scene sub-area from the scene data set, divides each scene data into K parts of scene image data by frame, establishes a three-dimensional coordinate system, and performs overlapping mapping on the scene image data with the same generation time node but different corresponding scene data types, thereby obtaining the multi-source image data of the corresponding scene sub-area at each time node within the data acquisition cycle, where K represents a natural number greater than 0;

[0019] Arrange the multi-source image data in time sequence, and perform pixel blurring processing on the optical image pixels of the optical image pixel layer of each multi-source pixel in each multi-source image data, and do not perform any operation on the image pixels of other layers;

[0020] Obtaining the magnitude of the pixel value difference ratio between adjacent optical image pixels in the optical image layer after pixel averaging, connecting two corresponding optical image pixels according to the magnitude of the pixel value difference ratio, and thereby obtaining the connection status between each optical image pixel in the optical image layer;

[0021] According to the connection status between each optical image pixel in the optical image layer, a plurality of optical scene objects are divided on the optical image layer, and the optical scene objects on the optical image layer are mapped onto the multi-source image data to obtain the optical image pixel composition status of the optical scene objects;

[0022] Then, according to the optical image pixel composition of the optical scene target, the corresponding multi-source pixels on the multi-source image data are connected, and then a number of scene targets are divided on the multi-source scene image.

[0023] Furthermore, the process of performing pixel blurring processing on the optical image pixels includes:

[0024] For any multi-source image data, its optical image pixels are segmented to obtain optical image layers corresponding to each multi-source image data. Starting from the first optical image pixel in the optical image layer, its pixel value is simultaneously accumulated with the pixel values ​​of each optical image pixel at an adjacent position, and then the average value is taken, and the average value is used as the pixel value corresponding to each optical image pixel, and so on, until the last optical image pixel in the optical image layer and each optical image pixel at an adjacent position are averaged.

[0025] Furthermore, the process of dividing the static scene objects and the dynamic scene objects includes:

[0026] According to the number of scene image data divisions of the scene data, K time nodes are set in the three-dimensional coordinate system, and then it is determined whether the multi-source pixels of each scene target in the multi-source image data at adjacent time nodes change in position;

[0027] If more than half of the multi-source pixels constituting the scene object change position, the scene object is judged to have shifted between the corresponding time nodes, otherwise it is judged that no shift has occurred;

[0028] Count the number of time nodes at which each scene target is displaced. If the number of time nodes at which the scene target is displaced is greater than or equal to 1 / 3K, mark the corresponding scene target as a dynamic scene target.

[0029] If the number of time nodes at which the scene target is displaced is less than 1 / 3K, the corresponding scene target is marked as a pre-static scene target;

[0030] The pre-static scene targets generated in different data cycles are matched with each other. If there is a match, the pre-static scene target is recorded as a static scene target. If there is no match, the pre-static scene target is saved and no operation is performed until a pre-static scene target is generated in a subsequent data cycle and there is a match with the pre-static scene target.

[0031] Furthermore, the scene target analysis module sequentially splices the static scene targets according to their distribution in the scene sub-areas, thereby obtaining a static scene image model. At the same time, a multi-source dynamic scene image model corresponding to each dynamic scene target is established based on the scene data set, and a body feature model of the corresponding scene personnel is generated based on the photos in the scene personnel identity data set, wherein the body feature model is composed of a facial feature model and a torso feature model of the scene personnel.

[0032] Furthermore, the process of marking the main identification points and the secondary identification points on the multi-source dynamic scene image model includes:

[0033] A primary identification point is set on the facial feature model, a plurality of primary identification points are set at the center of the torso feature model, and a plurality of secondary identification points are set at the edge of the torso feature model;

[0034] Matching the body feature model of each scene person with the multi-source dynamic scene image model. Since each multi-source dynamic scene image model is generated according to the scene data set of the application scene, it must contain the multi-source dynamic scene image model related to each scene person;

[0035] The body feature model matched with the multi-source dynamic scene image model is marked as a person multi-source image model, and the main identification points and the secondary identification points on the body feature model are annotated on the person multi-source image model, and the corresponding name and identity number are annotated on each person multi-source image model according to the scene person identity data set;

[0036] The multi-source dynamic scene image models that do not match the body feature model are marked as non-person multi-source image models.

[0037] Furthermore, the process of identifying the real-time scene target image model includes:

[0038] The target recognition module obtains the real-time scene data set of the application scene in the real-time data collection period from the scene data collection module, generates a number of real-time scene target image models according to the real-time scene data set, and divides the real-time scene target image model into a real-time static scene target image model and a real-time dynamic scene target image model by using a method of analyzing static scene targets and dynamic scene targets;

[0039] Matching each real-time static scene target image model with a static scene image model, and matching each real-time dynamic scene target image model with a non-personnel multi-source image model and a personnel multi-source image model;

[0040] Then, the identity information of the real-time dynamic scene target image model corresponding to the scene personnel is obtained based on the matching results.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] The present invention classifies scene targets into static scene targets and dynamic scene targets, constructs a static scene image model and several multi-source dynamic scene image models of the application scenario according to multi-source data of the static scene targets and the dynamic scene targets, sets multiple main recognition points and secondary recognition points on the multi-source dynamic scene image model, generates several real-time scene target image models according to the real-time scene data set, matches the real-time scene target image model with the multi-source dynamic scene image model and the static scene image model, and recognizes each real-time scene target image model according to the matching result, thereby improving the variety of scene object recognition while ensuring the accuracy of scene object recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention.

[0044] Figure 1 It is a schematic diagram of a scene image segmentation system based on local feature fusion of the present invention. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other implementation methods obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0046] like Figure 1 As shown, a scene image segmentation system based on local feature fusion includes a cloud computing platform, wherein the cloud computing platform is communicatively connected with a scene data acquisition module, a scene target segmentation module, a scene target analysis module, and a target recognition module;

[0047] The scene data acquisition module is used to set several types of sensors in the application scene and set several data acquisition cycles, and then call the sensors to collect scene data in each data acquisition cycle, and generate corresponding scene data sets to send to the scene target segmentation module;

[0048] The scene target segmentation module is used to construct each scene target in the application scene according to the scene data set, and extract multi-source data of each scene target from the scene data set, and then classify the scene targets into static scene targets and dynamic scene targets;

[0049] The scene target analysis module is used to construct a static scene image model and several multi-source dynamic scene image models of the application scene according to the multi-source data of the static scene target and the dynamic scene target, and set a plurality of primary identification points and secondary identification points on the multi-source dynamic scene image model;

[0050] The target recognition module is used to generate a plurality of real-time scene target image models according to the real-time scene data set, match the real-time scene target image models with multi-source dynamic scene image models and static scene image models, and recognize each real-time scene target image model according to the matching results.

[0051] Further, the working principle of the present invention is described below by way of examples:

[0052] The scene data acquisition module divides the target scene into N scene sub-areas of the same size, and sets numbers s1, s2, s3, ..., s N , where N represents a positive integer;

[0053] A plurality of sensors of the same type are arranged in each scene sub-area, and the types of the sensors include infrared sensors, cameras, etc. In this embodiment, the application scenarios include scenes such as housing communities and factories;

[0054] The same data collection cycle is set for all sensors. When the data collection cycle starts, the scene data collection module sends a data collection instruction to each sensor at the same time, and then each sensor collects various scene data of the scene sub-area where it is located according to the data collection instruction, wherein the scene data includes scene optical video data, scene infrared video data, etc.;

[0055] When the data collection cycle is over, each sensor uploads the collected data to the scene data collection module, and then the scene data collection module marks the corresponding scene sub-area number for each scene data, and integrates the scene data with the same number to generate a scene data set of the corresponding scene sub-area, and marks the corresponding scene sub-area number;

[0056] At the same time, community or factory managers upload residents' or workers' identity information to the scene data collection module. The resident or worker identity information includes the name, identity number and photo of the people entering and leaving the corresponding application scene. The scene data collection module then generates a scene personnel identity data set based on the resident or worker identity information.

[0057] Further, the scene data collection module sends all scene data sets and scene personnel identity data sets to the scene target analysis module;

[0058] The scene object segmentation module obtains the scene data of each scene sub-area from the scene data set, divides each scene data into K parts of scene image data by frame, establishes a three-dimensional coordinate system, and performs overlapping mapping on the scene image data with the same generation time node but different corresponding scene data types, thereby obtaining the multi-source image data of the corresponding scene sub-area at each time node within the data acquisition cycle, where K is a natural number greater than 0;

[0059] In this embodiment, each multi-source pixel in the multi-source image data is composed of pixels of multiple layers of image data, such as optical image pixels, infrared image pixels, etc.;

[0060] Arrange the multi-source image data in time sequence, and perform pixel blurring processing on the optical image pixels of the optical image pixel layer of each multi-source pixel in each multi-source image data, and do not perform any operation on the image pixels of other layers;

[0061] The process of performing pixel blurring processing on the optical image pixels includes:

[0062] For any multi-source image data, its optical image pixels are segmented to obtain optical image layers corresponding to each multi-source image data. Starting from the first optical image pixel in the optical image layer, its pixel value is simultaneously accumulated with the pixel values ​​of each optical image pixel at an adjacent position, and then the average value is taken, and the average value is used as the pixel value corresponding to each optical image pixel, and so on, until the last optical image pixel in the optical image layer is averaged with each optical image pixel at an adjacent position, wherein the pixel blurring process is used to provide data support for the subsequent extraction of several scene targets from the optical image pixel layer;

[0063] Determine whether the pixel value difference rate F between adjacent optical image pixels in the optical image layer after pixel averaging is within 5%, if so, connect the corresponding two optical image pixels, if not, do nothing;

[0064] It should be noted that the calculation formula of the pixel value difference rate F is: F = |(Yy)| / min(Y, y), where Y and y represent the pixel values ​​of two adjacent pixels;

[0065] Then, according to the connection status between each optical image pixel in the optical image layer, a plurality of optical scene objects are divided on the optical image layer.

[0066] Further, the optical scene target on the optical image layer is mapped onto the multi-source image data, and then the corresponding multi-source pixels on the multi-source image data are connected according to the optical image pixel composition of the optical scene target, and then a corresponding number of scene targets are divided on the multi-source scene image;

[0067] According to the number of scene image data divisions of the scene data, K time nodes are set on the three-dimensional coordinate system, and then it is determined whether the multi-source pixels of each scene target in the multi-source image data at adjacent time nodes change in position; wherein K represents a natural number greater than 0;

[0068] If more than half of the multi-source pixels constituting the scene object change position, the scene object is judged to have shifted between the corresponding time nodes, otherwise it is judged that no shift has occurred;

[0069] Count the number of time nodes at which each scene target is displaced. If the number of time nodes at which the scene target is displaced is greater than or equal to 1 / 3K, mark the corresponding scene target as a dynamic scene target.

[0070] If the number of time nodes at which the scene target is displaced is less than 1 / 3K, the corresponding scene target is marked as a pre-static scene target;

[0071] The pre-static scene targets generated in different data cycles are matched with each other. If 90% of the optical image pixels of the pre-static scene targets in two different data cycles are the same, it is judged that there is a match result, and the pre-static scene target is recorded as a static scene target. Otherwise, it is judged that there is no match result, and the pre-static scene target is saved and no operation is performed until a pre-static scene target is generated in a subsequent data cycle and there is a match result with the pre-static scene target.

[0072] Further, the scene object segmentation module sends the static scene objects and dynamic scene objects of each scene sub-area to the scene object analysis module;

[0073] The scene target analysis module sequentially splices each static scene target according to the distribution of the scene sub-area in which it is located, and then obtains a static scene image model. At the same time, it establishes a multi-source dynamic scene image model corresponding to each dynamic scene target based on the scene data set, and generates a physical feature model of the corresponding scene personnel based on the photos in the scene personnel identity data set;

[0074] In this embodiment, the multi-source dynamic scene image model is composed of multiple layers of dynamic image models, including an optical dynamic image model, an infrared dynamic image model, etc., which are stacked in sequence;

[0075] The body feature model is composed of a facial feature model and a torso feature model of the scene person;

[0076] A primary identification point is set on the facial feature model, a plurality of primary identification points are set at the center of the torso feature model, and a plurality of secondary identification points are set at the edge of the torso feature model;

[0077] In this embodiment, the primary identification points include the immutable parts such as the face and skeleton structure of the scene personnel, and the secondary identification points include the variable parts such as the clothing and handheld objects of the scene personnel;

[0078] Matching the body feature model of each scene person with the multi-source dynamic scene image model. Since each multi-source dynamic scene image model is generated according to the scene data set of the application scene, it must contain the multi-source dynamic scene image model related to each scene person;

[0079] The body feature model matched with the multi-source dynamic scene image model is marked as a person multi-source image model, and the main identification points and the secondary identification points on the body feature model are annotated on the person multi-source image model, and the corresponding name and identity number are annotated on each person multi-source image model according to the scene person identity data set;

[0080] The multi-source dynamic scene image models that do not match the body feature model are marked as non-person multi-source image models.

[0081] Further, the scene target analysis module sends all non-personnel multi-source image models, personnel multi-source image models and static scene image models to the target recognition module;

[0082] The target recognition module obtains the real-time scene data set of the application scene in the real-time data collection cycle from the scene data collection module;

[0083] Then the target recognition module generates a number of real-time scene target image models according to the real-time scene data set, and divides the real-time scene target image models into real-time static scene target image models and real-time dynamic scene target image models by using a method of analyzing static scene targets and dynamic scene targets;

[0084] Matching each real-time static scene target image model with a static scene image model, and matching each real-time dynamic scene target image model with a non-personnel multi-source image model and a personnel multi-source image model;

[0085] Then, the identity information of the real-time dynamic scene target image model corresponding to the scene personnel is obtained based on the matching results.

[0086] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A scene image segmentation system based on local feature fusion, comprising: A cloud computing platform, characterized in that the cloud computing platform is communicatively connected with a scene data acquisition module, a scene target segmentation module, a scene target analysis module and a target recognition module; The scene data acquisition module is used to set several types of sensors in the application scene, set several data acquisition cycles, and call the sensors to collect scene data in each data acquisition cycle; The scene target segmentation module is used to construct each scene target in the application scene according to the scene data set, and extract multi-source data of each scene target from the scene data set, and then classify the scene targets into static scene targets and dynamic scene targets; The process of building a scene target includes: Acquire various scene data of each scene sub-region from the scene data set, and divide each scene data into K parts of scene image data by frame; establish a three-dimensional coordinate system, overlap and map the scene image data with the same generation time node but different corresponding scene data types, and then obtain the multi-source image data of the corresponding scene sub-region at each time node within the data acquisition period; wherein K represents a natural number greater than 0; Arrange the multi-source image data in time sequence, and perform pixel blurring processing on the optical image pixels of the optical image pixel layer of each multi-source pixel in each multi-source image data, and do not perform any operation on the image pixels of other layers; Obtaining the magnitude of the pixel value difference ratio between adjacent optical image pixels in the optical image layer after pixel averaging, connecting two corresponding optical image pixels according to the magnitude of the pixel value difference ratio, and thereby obtaining the connection status between each optical image pixel in the optical image layer; According to the connection status between each optical image pixel in the optical image layer, a plurality of optical scene objects are divided on the optical image layer, and the optical scene objects on the optical image layer are mapped onto the multi-source image data to obtain the optical image pixel composition status of the optical scene objects; Then, according to the optical image pixel composition of the optical scene target, the corresponding multi-source pixels on the multi-source image data are connected, and then a plurality of scene targets are divided on the multi-source scene image; Perform pixel blurring on optical image pixels, including: For any multi-source image data, its optical image pixels are segmented to obtain optical image layers corresponding to each multi-source image data, and starting from the first optical image pixel in the optical image layer, its pixel value and the pixel values ​​of each optical image pixel at an adjacent position are simultaneously accumulated to obtain an average value, and the average value is used as the pixel value corresponding to each optical image pixel, and so on, until the last optical image pixel in the optical image layer and each optical image pixel at an adjacent position are averaged; The process of dividing static scene objects and dynamic scene objects includes: According to the number of scene image data divisions of the scene data, K time nodes are set in the three-dimensional coordinate system, and then it is determined whether the multi-source pixels of each scene target in the multi-source image data at adjacent time nodes have position changes; wherein K represents a natural number greater than 0; If more than half of the multi-source pixels constituting the scene object change position, the scene object is judged to have shifted between the corresponding time nodes, otherwise it is judged that no shift has occurred; Count the number of time nodes at which each scene target is displaced. If the number of time nodes at which the scene target is displaced is greater than or equal to 1 / 3K, the corresponding scene target is marked as a dynamic scene target; if the number of time nodes at which the scene target is displaced is less than 1 / 3K, the corresponding scene target is marked as a pre-static scene target. The pre-static scene targets generated in different data cycles are matched with each other. If there is a match, the pre-static scene target is recorded as a static scene target. If there is no match, the pre-static scene target is saved and no operation is performed until a pre-static scene target generated in a subsequent data cycle has a match result with the pre-static scene target; The scene target analysis module is used to construct a static scene image model and several multi-source dynamic scene image models of the application scene according to the multi-source data of the static scene target and the dynamic scene target, and set a plurality of primary identification points and secondary identification points on the multi-source dynamic scene image model; The target recognition module is used to generate a plurality of real-time scene target image models according to the real-time scene data set, match the real-time scene target image models with the multi-source dynamic scene image model and the static scene image model, and recognize each real-time scene target image model according to the matching results; The process of identifying the real-time scene target image model includes: The target recognition module obtains the real-time scene data set of the application scene in the real-time data collection period from the scene data collection module, generates a number of real-time scene target image models according to the real-time scene data set, and divides the real-time scene target image model into a real-time static scene target image model and a real-time dynamic scene target image model by using a method of analyzing static scene targets and dynamic scene targets; Each real-time static scene target image model is matched with a static scene image model, and each real-time dynamic scene target image model is matched with a non-personnel multi-source image model and a personnel multi-source image model, and then the identity information of the real-time dynamic scene target image model corresponding to the scene personnel is obtained according to the matching results.

2. A scene image segmentation system based on local feature fusion according to claim 1, characterized in that: The data collection process of the scene data collection module includes: The scene data acquisition module divides the target scene into N scene sub-areas of the same size, and sets a number for each scene sub-area, sets a plurality of sensors of the same type in each scene sub-area, and sets the same data acquisition cycle for all sensors. When the data acquisition cycle starts, the scene data acquisition module sends a data acquisition instruction to each sensor at the same time, and then each sensor collects various scene data of the scene sub-area where it is located according to the data acquisition instruction; wherein N represents a positive integer; When the data collection cycle is over, each sensor uploads the collected data to the scene data collection module, and then the scene data collection module marks the corresponding scene sub-area number for each scene data, and integrates the scene data with the same number to generate a scene data set of the corresponding scene sub-area, and marks the corresponding scene sub-area number; At the same time, community or factory managers upload residents' or workers' identity information to the scene data collection module. The resident or worker identity information includes the name, identity number and photo of the people entering and leaving the corresponding application scene. The scene data collection module then generates a scene personnel identity data set based on the resident or worker identity information.

3. The scene image segmentation system based on local feature fusion according to claim 2, characterized in that: The scene target analysis module is used to construct a static scene image model and several multi-source dynamic scene image models in the application scenario based on the multi-source data of static scene targets and dynamic scene targets, including: The scene target analysis module splices each static scene target in sequence according to the distribution of the scene sub-area in which it is located to obtain a static scene image model. At the same time, it establishes a multi-source dynamic scene image model corresponding to each dynamic scene target based on the scene data set, and generates a body feature model of the corresponding scene person based on the photos in the scene person identity data set. The body feature model consists of a facial feature model and a torso feature model of the scene person.

4. The scene image segmentation system based on local feature fusion according to claim 3, characterized in that: The process of labeling the main recognition points and the secondary recognition points on the multi-source dynamic scene image model includes: A primary identification point is set on the facial feature model, a plurality of primary identification points are set at the center of the torso feature model, and a plurality of secondary identification points are set at the edge of the torso feature model; The body feature model of each person in the scene is matched with the multi-source dynamic scene image model, and the body feature model that matches the multi-source dynamic scene image model is marked as a person multi-source image model, and the main identification points and secondary identification points on the body feature model are annotated on the person multi-source image model. In addition, the corresponding name and identity number are annotated for each person multi-source image model according to the scene person identity data set, and the multi-source dynamic scene image model that does not match the body feature model is marked as a non-person multi-source image model.

Images