Test place personnel management and control method based on deep learning

Abstract

The invention discloses a test place personnel management and control method based on deep learning. A data acquisition module, a personnel identity registration module, a workshop personnel identity recognition module, a dangerous place access people counting module and a detection result visualization module are included. The data acquisition module transmits the real-time video of the camera to the software front end for display; the personnel identity registration module performs face detection based on a RetinaFace improved model; the work cell personnel identity identification module is used for realizing identification of personnel identities in a work cell on the basis of a RetinaFace improved model and an Insight Face network; the dangerous place entering and exiting people counting module counts the number of people entering and exiting the dangerous place in real time based on a YOLOv5 target detection algorithm and a DeepSort multi-people tracking algorithm; and the detection result visualization module is used for displaying the video detection result processed by the deep learning algorithm in real time in a software front-end interface. The method has the advantage of light weight, can be deployed at a mobile terminal or an embedded terminal, and widens the application range of the algorithm while reducing the hardware cost.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning technology, and specifically relates to a personnel management method for a test site based on deep learning. Background Art

[0002] In the process of weapon system development, tests are often conducted in dangerous test sites, where the environmental conditions and operating procedures are special. In dangerous test sites, there may be extreme environmental conditions such as high temperature, high pressure, high radiation, and high corrosion. Considering the safety and confidentiality of the test, it is impossible to effectively track and detect personnel in the test site, and there are great risks to the personal safety of on-site test personnel.

[0003] In addition to on-site tests, some personnel are also required to monitor and control the data of the on-site tests in the workshop. In order to prevent unrelated personnel from entering the workshop, the identities of the personnel in the workshop need to be identified.

[0004] At present, the personnel management measures in dangerous test scenarios are only constrained by some rules and regulations and test processes, which rely largely on human consciousness and cannot be managed scientifically and effectively. Therefore, there is an urgent need for an intelligent personnel management method to identify and track test personnel. Summary of the invention

[0005] In order to overcome the shortcomings of the prior art, the present invention provides a method for personnel management and control of a test site based on deep learning, including a data acquisition module, a personnel identity registration module, a workshop personnel identity recognition module, a module for counting the number of people entering and leaving dangerous places, and a detection result visualization module. The data acquisition module transmits the real-time video of the camera to the software front-end for display; the personnel identity registration module performs face detection based on the RetinaFace improved model; the workshop personnel identity recognition module realizes the identification of the personnel identity of the workshop based on the RetinaFace improved model and the InsightFace network; the module for counting the number of people entering and leaving dangerous places is based on the YOLOv5 target detection algorithm and the DeepSort multi-personnel tracking algorithm, and counts the number of people entering and leaving dangerous places in real time; the detection result visualization module is used to display the video detection results processed in real time by the deep learning algorithm in the software front-end interface. The method of the present invention has the advantage of being lightweight and can be deployed on a mobile terminal or an embedded terminal, which reduces the hardware cost while expanding the application scope of the algorithm.

[0006] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0007] Step 1: Build a deep learning-based test site personnel management and control system, including a data collection module, a personnel identity registration module, a workshop personnel identity recognition module, a number of people entering and leaving dangerous places, and a detection result visualization module;

[0008] Step 2: The data acquisition module includes a personnel identification camera in the workshop and a personnel tracking camera in the dangerous test site, which is used to capture the personnel information in the scene in real time and provide it to the software for subsequent algorithm processing;

[0009] Step 3: The personnel identity registration module process is as follows:

[0010] Step 3-1: Read the real-time video stream of the data acquisition module;

[0011] Step 3-2: Use the RetinaFace improved model to detect faces in the video stream;

[0012] Step 3-3: Filter out the face with the largest corresponding face frame in the real-time video stream;

[0013] Step 3-4: Click the "Collect Face" button on the interface to capture the face and fill in the corresponding identity information;

[0014] Step 3-5: Click “Save” in the interface to save the face and corresponding identity information into the database.

[0015] Step 4: The workflow of the workshop personnel identification module is as follows:

[0016] Step 4-1: Read the real-time video stream of the data acquisition module;

[0017] Step 4-2: Click "Entry Area Selection" and "Exit Area Selection" in the interface to specify the entry area and exit area for detecting personnel identities respectively;

[0018] Step 4-3: Use the RetinaFace improved model to detect faces in the specified area;

[0019] Step 4-4: The RetinaFace improved model obtains the face detection frame information and face key point information, and sends this face information to the InsightFace face recognition algorithm for comparison with the face images in the face database established by the face registration module to obtain the person identification result. If the face match is successful, the real-time person information is displayed on the front-end interface, otherwise, the stranger face image is captured and displayed in the interface;

[0020] Step 5: The process of the number of people entering and leaving dangerous places is as follows:

[0021] Step 5-1: Read the real-time video stream of the data acquisition module;

[0022] Step 5-2: Click "Set Goal Line" in the interface to set the goal line;

[0023] Step 5-3: Use the YOLOv5 target detection model to detect people in the video;

[0024] Step 5-4: Use DeepSort multi-target tracking algorithm to achieve multi-person tracking;

[0025] Step 5-5: When the center of the tracking frame passes through the door line set in step 5-2, the direction of movement is determined. If it is the direction of entering the door, the number of people entering the door is increased by 1, otherwise the number of people leaving the door is increased by 1.

[0026] Step 6: The detection result visualization module is used to display the video detection results processed by the deep learning algorithm in real time in the software front-end interface.

[0027] Preferably, the improved RetinaFace model uses MobileNetv2 instead of ResNet50 as the backbone network of RetinaFace.

[0028] Preferably, the MobileNetv2 uses a depthwise separable convolutional structure; MobileNetv2 introduces an inverted residual structure and a linear bottleneck layer; the input of the inverted residual structure is a low-dimensional feature, which is then filtered and then output in a reduced dimension; the features of the input image are amplified at the previous expansion layer of the linear bottleneck layer, and then high-dimensional features are extracted through a deep convolution unit, and finally compressed through a mapping layer.

[0029] Preferably, the formula of the loss function of the InsightFace face recognition algorithm is as follows:

[0030]

[0031] Among them, Larc represents the loss value, N represents the number of samples; s is the scale parameter, which is used to adjust the radius of the hypersphere where the facial image features are located; Representation parameters and the face image feature vector x i , m is used to control the distance between different face image categories, and n represents the number of face categories in the dataset.

[0032] Preferably, the network structure of the YOLOv5 target detection model consists of three parts: BackBone, Neck and Head; wherein BackBone is the backbone network, and its function is to extract the features of the input image; the Neck layer fuses the multi-layer feature maps from BackBone and extracts higher-level semantic information to improve the performance of target detection; the Head layer is the last layer of the target detection network, which is used to predict the category and position of the target, receives the feature map from the Neck layer, and generates the category and position prediction of the target box according to each feature point.

[0033] Preferably, the DeepSort multi-target tracking algorithm first predicts the next frame position of the target person through the Kalman filter algorithm, then performs cascade matching, and finally uses the Hungarian algorithm for data association.

[0034] Preferably, the front-end interface is developed based on PyQt5.

[0035] The beneficial effects of the present invention are as follows:

[0036] 1) This paper proposes an InsightFace face recognition method based on the improved RetinaFace model. The improved RetinaFace model has the advantage of being lightweight and can be deployed on mobile or embedded terminals, which reduces hardware costs and expands the application scope of the algorithm.

[0037] 2) The present invention proposes a DeepSort multi-person tracking method based on the YOLOv5 target detection model, which can meet the performance requirements of the multi-person tracking model in the experimental scenario.

[0038] 3) The present invention designs a hardware platform and software system for personnel identification and entry and exit counting in dangerous test scenarios. A hardware platform suitable for actual needs is built, and an intelligent personnel management and control software system for dangerous test scenarios is designed and implemented based on the PyQt framework, including a personnel identity registration module, a workshop personnel identity recognition module, and a module for counting the number of people entering and leaving dangerous places. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Develop a framework for the system of the present invention.

[0040] Figure 2 This is a flow chart of the face registration procedure of the method of the present invention.

[0041] Figure 3 This is the MobileNetv2 network structure of the method of the present invention.

[0042] Figure 4 The present invention is a flowchart of the process of identifying the identity of personnel in a workshop according to the method of the present invention.

[0043] Figure 5 This is the interface for counting the number of people entering and leaving the dangerous area according to the method of the present invention.

[0044] Figure 6 The present invention is a flowchart of the procedure for counting the number of people entering and leaving a dangerous area.

[0045] Figure 7 It is the hardware platform of the personnel management and control system of the method of the present invention.

[0046] Figure 8 This is a schematic diagram of the installation of a camera for identifying the identity of personnel in a workshop according to the method of the present invention. DETAILED DESCRIPTION

[0047] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0048] The purpose of the present invention is to provide a test site personnel control system based on deep learning, which uses theories such as deep learning and machine vision as means to achieve intelligent and automated personnel control in dangerous test sites.

[0049] In order to achieve the above object, the technical solution adopted by the present invention is:

[0050] Build a deep learning-based test site personnel management and control system, including a data collection module, a personnel identity registration module, a workshop personnel identity recognition module, a number of people entering and leaving dangerous places, and a test result visualization module;

[0051] The personnel identity registration module includes the following steps:

[0052] Step 1) reading the real-time video stream of the data acquisition module;

[0053] Step 2) Use the RetinaFace improved model to detect faces in the video stream;

[0054] Step 3) Filter out the face with the largest corresponding face frame in the real-time video stream;

[0055] Step 4) Click the "Collect Face" button on the interface to capture the face and fill in the corresponding identity information;

[0056] Step 5) Click “Save” in the interface to save the face and corresponding identity information into the database.

[0057] In step 2), the RetinaFace improved model uses MobileNetv2 instead of ResNet50 as the backbone network of RetinaFace.

[0058] MobileNetv2 uses a deep separable convolution structure. Compared with the ordinary convolution structure, this structure reduces the amount of calculation of the network by several times while having the same number of weight parameters and has a good detection effect. In addition, MobileNetv2 introduces an inverted residual structure and a linear bottleneck layer. The difference between the inverted residual structure and the normal residual structure is that it inputs low-dimensional features, then performs dimensionality up-filtering, and then performs dimensionality down-output. The above approach is achieved through a linear bottleneck layer. The features of the input image are amplified at the previous expansion layer of the linear bottleneck layer, and then high-dimensional features are extracted through a deep convolution unit. Finally, they are compressed through a mapping layer. The above processing steps can make the model lighter and have a faster reasoning speed.

[0059] The workshop personnel identification module comprises the following steps:

[0060] Step 1) reading the real-time video stream of the data acquisition module;

[0061] Step 2) Click "Entry Area Selection" and "Exit Area Selection" in the interface to specify the entry area and exit area for detecting personnel identities respectively;

[0062] Step 3) Use the RetinaFace improved model to detect faces in the specified area;

[0063] Step 4) The face detection frame information and face key point information obtained by the RetinaFace improved model are sent to the InsightFace algorithm for comparison with the face images in the face database established by the face registration module to obtain the person identity recognition result. If the face match is successful, the real-time person information is displayed on the front-end interface, otherwise the stranger face image is captured and displayed in the interface.

[0064] The InsightFace face recognition algorithm in step 4) optimizes the face recognition loss function by improving the feature vector normalization and additive angle interval, thereby improving the inter-class separability while strengthening the intra-class compactness and inter-class differences. The formula of its loss function is as follows:

[0065]

[0066] Larc represents the loss value, N represents the number of samples, and s is the scale parameter used to adjust the radius of the hypersphere where the facial image features are located. Representation parameters and the face image feature vector x i , m is used to control the distance between different face image categories, and n represents the number of face categories in the dataset.

[0067] The module for counting people entering and leaving the dangerous area comprises the following steps:

[0068] Step 1) reading the real-time video stream of the data acquisition module;

[0069] Step 2) Click "Set Goal Line" in the interface to set the goal line;

[0070] Step 3) Use the YOLOv5 model to detect people in the video;

[0071] Step 4) Use DeepSort algorithm to achieve multi-person tracking;

[0072] Step 5) When the center of the tracking frame passes through the door line set in step 2, the direction of movement is determined. If it is the direction of entering the door, the number of people entering the door is increased by one, otherwise the number of people leaving the door is increased by one.

[0073] The network structure of the YOLOv5 target detection model in step 3) consists of three parts: BackBone, Neck and Head. Among them, BackBone is the backbone network, and its main function is to extract the features of the input image; the Neck layer is designed to fuse the multi-layer feature maps from BackBone and extract higher-level semantic information to improve the performance of target detection; the Head layer is the last layer of the target detection network, which is used to predict the category and position of the target. It receives the feature map from the Neck layer and generates the category and position prediction of the target box according to each feature point.

[0074] In the process of DeepSort multi-target tracking algorithm in step 4), the next frame position of the target person is first predicted by Kalman filter algorithm, then cascade matching is performed, and finally the Hungarian algorithm is used for data association.

[0075] The steps of the DeepSort multi-target tracking algorithm based on the YOLOv5 model include:

[0076] Step 1) Use the YOLOv5 model to detect the image of the input video to obtain the detection frames of multiple targets;

[0077] Step 2) Initializing the Kalman filter using the features of multiple targets extracted by the convolutional neural network;

[0078] Step 3) The Kalman filter predicts the position of the target in the current frame based on the information of the target detection bounding box in the previous frame, thereby outputting the target prediction bounding box;

[0079] Step 4) Calculate the ratio of the intersection and union of the areas of the target detection bounding box and the predicted bounding box, and then perform cascade matching;

[0080] Step 5) Use the Hungarian algorithm to perform data association to obtain the best match. If the match is successful, the target coordinates are output. Otherwise, a temporary tracker is assigned to the target, and if the next three consecutive frames are matched successfully, the target is set as a new target; if the next three consecutive frames fail to match, the target is considered to be temporarily lost and the target information is saved. If there is no corresponding trajectory matching the target in the next 60 frames, the tracker of the target is deleted;

[0081] Step 6) Repeat steps 3 to 5 until all input video frames are processed.

[0082] Based on the above modules, the present invention designs a corresponding personnel management and control hardware platform based on the actual application site, and the hardware platform consists of a visual module and a data operation module. The collaborative work of these hardware devices can improve the efficiency and accuracy of personnel identification and tracking.

[0083] The visual module mainly includes personnel identification cameras in the workshop and personnel tracking cameras in the dangerous test site. These cameras are used to capture personnel information in the scene in real time and provide it to the software for subsequent algorithm processing.

[0084] The data computing module is the core component of the platform. It uses a high-performance server and is equipped with hardware devices such as CPU, GPU, memory and hard disk, which can achieve efficient image processing and data calculation.

[0085] Example:

[0086] See also Figure 1 The present invention provides a test site personnel management and control system based on deep learning, including a data acquisition module, a face registration module, a workshop personnel detection and recognition module, a number of people entering and leaving dangerous places counting module and a detection result visualization module.

[0087] See also Figure 2 The personnel identity registration module comprises the following steps:

[0088] Step 1) reading the real-time video stream of the data acquisition module;

[0089] Step 2) Use the RetinaFace improved model to detect faces in the video stream;

[0090] The improved RetinaFace model uses MobileNetv2 instead of ResNet50 as the backbone network of RetinaFace.

[0091] See also Figure 3MobileNetv2 uses a deep separable convolution structure. Compared with the ordinary convolution structure, this structure reduces the amount of calculation of the network by several times while having the same number of weight parameters and has a good detection effect. In addition, MobileNetv2 introduces an inverted residual structure and a linear bottleneck layer. The difference between the inverted residual structure and the normal residual structure is that it inputs low-dimensional features, then performs dimensionality up-filtering, and then performs dimensionality down-output. The above approach is achieved through a linear bottleneck layer. The features of the input image are amplified at the previous expansion layer of the linear bottleneck layer, and then high-dimensional features are extracted through a deep convolution unit, and finally compressed through a mapping layer. The above processing steps can make the model lighter and have a faster reasoning speed.

[0092] Step 3) Filter out the face with the largest corresponding face frame in the real-time video stream;

[0093] Step 4) Click the "Collect Face" button on the interface to capture the face and fill in the corresponding identity information;

[0094] Step 5) Click “Save” in the interface to save the face and corresponding identity information into the database.

[0095] See also Figure 4 The workshop personnel identification module comprises the following steps:

[0096] Step 1) reading the real-time video stream of the data acquisition module;

[0097] Step 2) Click "Entry Area Selection" and "Exit Area Selection" in the interface to specify the entry area and exit area for detecting personnel identities respectively;

[0098] Step 3) Use the RetinaFace improved model to detect faces in the specified area;

[0099] Step 4) The face detection frame information and face key point information obtained by the RetinaFace improved model are sent to the InsightFace algorithm for comparison with the face images in the face database established by the face registration module to obtain the person identity recognition result. If the face match is successful, the real-time person information is displayed on the front-end interface, otherwise the stranger face image is captured and displayed in the interface.

[0100] The InsightFace face recognition algorithm optimizes the face recognition loss function. By improving the feature vector normalization and additive angle interval, it improves the inter-class separability while strengthening the intra-class compactness and inter-class differences. The formula of its loss function is as follows:

[0101]

[0102] See also Figure 5 and Figure 6 The module for counting people entering and leaving the dangerous area comprises the following steps:

[0103] Step 1) reading the real-time video stream of the data acquisition module;

[0104] Step 2) Click "Set Goal Line" in the interface to set the goal line;

[0105] Step 3) Use the YOLOv5 model to detect people in the video;

[0106] The network structure of the YOLOv5 target detection model consists of three parts: BackBone, Neck, and Head. Among them, BackBone is the backbone network, and its main function is to extract the features of the input image; the Neck layer is designed to fuse the multi-layer feature maps from BackBone and extract higher-level semantic information to improve the performance of target detection; the Head layer is the last layer of the target detection network, which is used to predict the category and position of the target. It receives the feature map from the Neck layer and generates the category and position prediction of the target box based on each feature point.

[0107] Step 4) Use DeepSortt multi-target tracking algorithm to achieve multi-person tracking;

[0108] During the tracking process of the DeepSort multi-target tracking algorithm, the Kalman filter algorithm is first used to predict the next frame position of the target person, then cascade matching is performed, and finally the Hungarian algorithm is used for data association.

[0109] Step 5) When the center of the tracking frame passes through the door line set in step 2, the direction of movement is determined. If it is the direction of entering the door, the number of people entering the door is increased by one, otherwise the number of people leaving the door is increased by one.

[0110] The steps of the DeepSort multi-target tracking algorithm based on the YOLOv5 model include:

[0111] Step 1) Use the YOLOv5 model to detect the image of the input video to obtain the detection frames of multiple targets;

[0112] Step 2) Initializing the Kalman filter using the features of multiple targets extracted by the convolutional neural network;

[0113] Step 3) The Kalman filter predicts the position of the target in the current frame based on the information of the target detection bounding box in the previous frame, thereby outputting the target prediction bounding box;

[0114] Step 4) Calculate the ratio of the intersection and union of the areas of the target detection bounding box and the predicted bounding box, and then perform cascade matching;

[0115] Step 5) Use the Hungarian algorithm to perform data association to obtain the best match. If the match is successful, the target coordinates are output. Otherwise, a temporary tracker is assigned to the target, and if the next three consecutive frames are matched successfully, the target is set as a new target; if the next three consecutive frames fail to match, the target is considered to be temporarily lost and the target information is saved. If there is no corresponding trajectory matching the target in the next 60 frames, the tracker of the target is deleted;

[0116] Step 6) Repeat steps 3) to 5) until all input video frames are processed.

[0117] See also Figure 7 Based on the above modules, the present invention designs a corresponding personnel management and control hardware platform based on the actual application site. The hardware platform consists of a visual module and a data operation module. The hardware platform parameter configuration is shown in Table 1. The collaborative work of these hardware devices can improve the efficiency and accuracy of personnel identification and tracking.

[0118] Table 1 Hardware platform parameter configuration

[0119]

[0120] The visual module mainly includes personnel identification cameras in the workshop and personnel tracking cameras in the dangerous test site. These cameras are used to capture personnel information in the scene in real time and provide it to the software for subsequent algorithm processing.

[0121] The data computing module is the core component of the platform. It uses a high-performance server and is equipped with hardware devices such as CPU, GPU, memory and hard disk, which can achieve efficient image processing and data calculation.

[0122] In order to detect and identify the identity of a specific person and determine whether he / she has entered the workshop, the workshop personnel identification system needs to place two cameras with a field of view of 102 degrees at the door of the workshop. The larger field of view can ensure that people in different directions can be detected in the widest possible range. The camera is placed on a tripod, facing indoors and outdoors respectively, to ensure that the front face image of the relevant person can be clearly captured whether entering or leaving the workshop to accurately confirm his / her identity. Figure 8 .

Claims

1. A method for personnel management and control in a test site based on deep learning, characterized in that: The steps include: Step 1: Build a deep learning-based test site personnel management and control system, including a data collection module, a personnel identity registration module, a workshop personnel identity recognition module, a number of people entering and leaving dangerous places, and a detection result visualization module; Step 2: The data acquisition module includes a personnel identification camera in the workshop and a personnel tracking camera in the dangerous test site, which is used to capture the personnel information in the scene in real time and provide it to the software for subsequent algorithm processing; Step 3: The personnel identity registration module process is as follows: Step 3-1: Read the real-time video stream of the data acquisition module; Step 3-2: Use the RetinaFace improved model to detect faces in the video stream; Step 3-3: Filter out the face with the largest corresponding face frame in the real-time video stream; Step 3-4: Click the "Collect Face" button on the interface to capture the face and fill in the corresponding identity information; Step 3-5: Click "Save" in the interface to save the face and corresponding identity information into the database; Step 4: The workflow of the workshop personnel identification module is as follows: Step 4-1: Read the real-time video stream of the data acquisition module; Step 4-2: Click "Entry Area Selection" and "Exit Area Selection" in the interface to specify the entry area and exit area for detecting personnel identities respectively; Step 4-3: Use the RetinaFace improved model to detect faces in the specified area; Step 4-4: The RetinaFace improved model obtains the face detection frame information and face key point information, and sends this face information to the InsightFace face recognition algorithm for comparison with the face images in the face database established by the face registration module to obtain the person identification result. If the face match is successful, the real-time person information is displayed on the front-end interface, otherwise, the stranger face image is captured and displayed in the interface; Step 5: The process of the number of people entering and leaving dangerous places is as follows: Step 5-1: Read the real-time video stream of the data acquisition module; Step 5-2: Click "Set Goal Line" on the interface to set the goal line; Step 5-3: Use the YOLOv5 target detection model to detect people in the video; Step 5-4: Use DeepSort multi-target tracking algorithm to achieve multi-person tracking; Step 5-5: When the center of the tracking frame passes through the door line set in step 5-2, the direction of movement is determined. If it is the direction of entering the door, the number of people entering the door is increased by 1, otherwise the number of people leaving the door is increased by 1; Step 6: The detection result visualization module is used to display the video detection results processed by the deep learning algorithm in real time in the software front-end interface.

2. A method for controlling personnel in a test site based on deep learning according to claim 1, characterized in that: The improved RetinaFace model uses MobileNetv2 instead of ResNet50 as the backbone network of RetinaFace.

3. A method for controlling personnel in a test site based on deep learning according to claim 2, characterized in that: The MobileNetv2 uses a depthwise separable convolutional structure; MobileNetv2 introduces an inverted residual structure and a linear bottleneck layer; the input of the inverted residual structure is a low-dimensional feature, which is then filtered and then output in a reduced dimension; the features of the input image are amplified at the previous expansion layer of the linear bottleneck layer, and then high-dimensional features are extracted through a deep convolution unit, and finally compressed through a mapping layer.

4. The method for controlling personnel in a test site based on deep learning according to claim 1, characterized in that: The formula of the loss function of the InsightFace face recognition algorithm is as follows: Among them, Larc represents the loss value, N represents the number of samples; s is the scale parameter, which is used to adjust the radius of the hypersphere where the facial image features are located; Representation parameters and the face image feature vector x i , m is used to control the distance between different face image categories, and n represents the number of face categories in the dataset.

5. The method for controlling personnel in a test site based on deep learning according to claim 1, characterized in that: The network structure of the YOLOv5 target detection model consists of three parts: BackBone, Neck and Head. Among them, BackBone is the backbone network, which is used to extract the features of the input image. The Neck layer fuses the multi-layer feature maps from BackBone and extracts higher-level semantic information to improve the performance of target detection. The Head layer is the last layer of the target detection network, which is used to predict the category and position of the target, receives the feature map from the Neck layer, and generates the category and position prediction of the target box according to each feature point.

6. The method for controlling personnel in a test site based on deep learning according to claim 1, characterized in that: The DeepSort multi-target tracking algorithm first predicts the next frame position of the target person through the Kalman filter algorithm, then performs cascade matching, and finally uses the Hungarian algorithm for data association.

7. The method for controlling personnel in a test site based on deep learning according to claim 1, characterized in that: The front-end interface is developed based on PyQt5.

Images