Security detection method and system

Abstract

The invention discloses a safety detection method and system, and belongs to the field of safety detection, through the steps of identity recognition, position positioning, behavior analysis and abnormal event and dangerous event alarm, dangerous events (falling down, falling into water and the like) and abnormal event detection (crowds, approaching a forbidden zone / dangerous zone and the like) are detected in real time, and the potential safety hazards of old people are reduced. Through a position coding technology, non-inductive positioning of the elderly in the park is realized by adopting a monitoring video, and meanwhile, gait features and facial features of the detected pedestrian are extracted and fused, so that the accuracy and high applicability of identity recognition of the elderly can be remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of safety detection, and in particular to a safety detection method and system. Background Art

[0002] The global population is rapidly aging, and it is estimated that by 2050, the total number of people aged 60 and above will reach 2 billion. my country has the largest elderly population in the world, and the aging level continues to deepen. Aging-related degenerative diseases, neurological diseases and cognitive dysfunction induce behavioral safety problems in the elderly. With the development of social economy and changes in family structure, the phenomenon of empty nest and low birth rate is obvious. More and more families find it difficult to take care of the elderly alone, and the demand for institutional care is increasing.

[0003] It has become particularly important to ensure the behavioral safety of the elderly within the institutional campus, to achieve accurate indoor and outdoor positioning of the elderly, and to promptly detect risks such as falls, falling into water, and approaching dangerous areas. This is a prerequisite for providing high-quality health care services for the elderly. Summary of the invention

[0004] In order to overcome the deficiencies of the prior art, one of the objectives of the present invention is to provide a safety detection method that can timely detect dangerous risks of the elderly.

[0005] In order to overcome the deficiencies of the prior art, a second object of the present invention is to provide an elderly safety detection system that can timely detect dangerous risks of the elderly.

[0006] One of the purposes of the present invention is achieved by the following technical solution:

[0007] A safety detection method comprises the following steps:

[0008] Identity recognition: extract gait and facial multimodal feature vectors through surveillance video, perform multimodal biometric recognition, match the recognition results with the human biometric database, and realize human identity recognition;

[0009] Position positioning: Divide the area where the person is located into multiple physical subspaces, each subspace is equipped with a surveillance camera, and establish a video-camera-subspace association relationship according to the regional location coding strategy, and obtain the location information of the person with the corresponding identity based on the identity recognition result;

[0010] Behavior analysis: Using surveillance video as input data for human behavior analysis, an instance segmentation model is used to obtain information on human body boundaries, external environment boundaries, risk factor boundaries, and the number of elderly pedestrians. A video action understanding model is used to identify human action states and action durations. Human behavior events are judged based on the location information, action states, action durations, external environment boundaries, risk factor boundaries, and the number of elderly pedestrians.

[0011] Abnormal event and dangerous event alarm: When a person's behavior is dangerous or abnormal, an alarm will be issued and the person's identity information and corresponding location information will be output.

[0012] Furthermore, in the position positioning step, the area where the person is located is divided into multiple physical subspaces, specifically: the area is divided into at least two of an activity area, a fenced area, a work restricted area, a dangerous area, a water area, and a staircase.

[0013] Furthermore, in the position positioning step, according to the regional position coding strategy, a video-camera-subspace association relationship is established, specifically: multiple cameras are encoded separately, and each camera is multi-level encoded, and the multi-level encoding includes at least two of the security level, area name, sub-area name, and detailed location.

[0014] Furthermore, in the behavior analysis step, the action states of a person identified by using a video action understanding model include walking, sitting, standing, getting up, waving arms, squatting, climbing stairs, and lying down.

[0015] Furthermore, in the behavior analysis step, the dangerous and abnormal situations occurring in the behavior events include falling, falling into water, gathering, and approaching dangerous areas.

[0016] Furthermore, the identity recognition step specifically includes:

[0017] Video detection: Obtain camera surveillance video, use the pedestrian detection model to infer each image frame of the surveillance video stream, and obtain the confidence score of the existence of pedestrians and the confidence score of the existence of faces in the image frame. When the confidence score of pedestrians and / or the confidence score of faces is greater than the threshold, it is determined that pedestrians and / or faces are detected;

[0018] Gait feature extraction: When a pedestrian is detected, the effective image area where the pedestrian exists is cropped, and a pedestrian image sequence frame group is constructed. The cropped pedestrian image area is aligned, and the pedestrian image sequence is judged using the behavior recognition model to obtain the action state of the pedestrian in each frame image. The pedestrian image sequence in the "walking" action state is selected, and the pedestrian gait features are obtained based on the pedestrian image sequence in the "walking" action state;

[0019] Facial feature extraction: When a face is detected, the effective image area where the face exists is cropped, the face image is enhanced, the face image is diffused in multiple dimensions, the data fusion of multi-angle face information is completed, and the facial features of the fused image are extracted;

[0020] Feature fusion: Calculate the gait feature significance and the facial feature significance, determine the feature fusion weighting method according to the numerical values ​​of the gait feature significance and the facial feature significance, and fuse the gait features and facial features according to the determined feature fusion weighting method to obtain a feature fusion vector;

[0021] Establishing a human biometric identity database: Collecting each person's identity information and biological information to establish a human biometric identity database, the biological information includes facial images and gait information;

[0022] Identify the identity information of the person to be tested: match the biometric fusion features of the person to be tested with the biometric features in the identity database to identify the identity information of the person to be tested.

[0023] Furthermore, in the gait feature extraction step, the construction of the pedestrian image sequence frame group is specifically: using a pedestrian detection frame based on pedestrian detection model reasoning, giving a boundary filling pixel size, cropping the pedestrian image area, and constructing a pedestrian image sequence frame group under a continuous monitoring video stream.

[0024] Furthermore, in the gait feature extraction step, the image alignment operation performed on the cropped pedestrian image region is specifically performed by using feature matching and pixel interpolation to perform the image alignment operation.

[0025] Furthermore, in the feature fusion step, the gait feature prominence and the facial feature prominence are calculated based on the corresponding feature approximation, model confidence, image quality score and feature area coverage, respectively.

[0026] The second object of the present invention is achieved by adopting the following technical solution:

[0027] A safety detection system, used to implement any one of the above safety detection methods, the safety detection system comprising

[0028] The acquisition module, multiple cameras located in different areas collect surveillance videos of the corresponding areas;

[0029] An analysis module, which analyzes the identity information of the person in the surveillance video, obtains the location information of the person according to the coding of the camera, and performs behavior analysis according to the location information and the surveillance video;

[0030] The output module will sound an alarm when a person's behavior is dangerous or abnormal, and output the person's identity information and corresponding location information.

[0031] Compared with the existing technology, the safety detection method of the present invention can realize the imperceptible positioning of the elderly in the park through monitoring video, and at the same time extract and integrate the gait characteristics and facial features of the detected pedestrians, which can significantly improve the accuracy and high applicability of the identity recognition of the elderly, and perform real-time detection of dangerous events (falling, falling into water, etc.) and abnormal event detection (gathering, approaching restricted / dangerous areas, etc.), thereby reducing safety hazards for the elderly. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a flow chart of the safety detection method of the present invention;

[0033] Figure 2 for Figure 1 Flow chart of identity recognition in the security detection method;

[0034] Figure 3 for Figure 2 Flowchart of gait feature extraction in the identification step;

[0035] Figure 4 for Figure 2 Flowchart of facial feature extraction in the identification step;

[0036] Figure 5 for Figure 2 Schematic diagram of a human biometric identity database in the identification step;

[0037] Figure 6 for Figure 1 Schematic diagram of subspace division in the safety detection method;

[0038] Figure 7 for Figure 1 Schematic diagram of behavioral analysis in the security detection method. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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 embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0041] like Figure 1As shown, the present invention provides a safety detection method, comprising the following steps:

[0042] Identity recognition: extract gait and facial multimodal feature vectors through surveillance video, perform multimodal biometric recognition, match the recognition results with the human biometric database, and realize human identity recognition;

[0043] Position positioning: Divide the area where the person is located into multiple physical subspaces, each subspace is equipped with a surveillance camera, and establish a video-camera-subspace association relationship according to the regional location coding strategy, and obtain the location information of the person with the corresponding identity based on the identity recognition result;

[0044] Behavior analysis: Using surveillance video as input data for human behavior analysis, an instance segmentation model is used to obtain information on human body boundaries, external environment boundaries, risk factor boundaries, and the number of elderly pedestrians. A video action understanding model is used to identify human action states and action durations. Human behavior events are judged based on the location information, action states, action durations, external environment boundaries, risk factor boundaries, and the number of elderly pedestrians.

[0045] Abnormal event and dangerous event alarm: When a person's behavior is dangerous or abnormal, an alarm will be issued and the person's identity information and corresponding location information will be output.

[0046] Please continue reading Figure 2 , identity identification specifically includes the following steps:

[0047] Video detection: Obtain camera surveillance video, use the pedestrian detection model to infer each image frame of the surveillance video stream, and obtain the confidence score of the existence of pedestrians and the confidence score of the existence of faces in the image frame. When the confidence score of pedestrians and / or the confidence score of faces is greater than the threshold, it is determined that pedestrians and / or faces are detected;

[0048] Gait feature extraction: When a pedestrian is detected, the effective image area where the pedestrian exists is cropped, and a pedestrian image sequence frame group is constructed. The cropped pedestrian image area is aligned, and the pedestrian image sequence is judged using the behavior recognition model to obtain the action state of the pedestrian in each frame image. The pedestrian image sequence in the "walking" action state is selected, and the pedestrian gait features are obtained based on the pedestrian image sequence in the "walking" action state;

[0049] Facial feature extraction: When a face is detected, the effective image area where the face exists is cropped, the face image is enhanced, the face image is diffused in multiple dimensions, the data fusion of multi-angle face information is completed, and the facial features of the fused image are extracted;

[0050] Feature fusion: Calculate the gait feature significance and the facial feature significance, determine the feature fusion weighting method according to the numerical values ​​of the gait feature significance and the facial feature significance, and fuse the gait features and facial features according to the determined feature fusion weighting method to obtain a feature fusion vector;

[0051] Establishing a human biometric identity database: Collecting each person's identity information and biological information to establish a human biometric identity database, the biological information includes facial images and gait information;

[0052] Identify the identity information of the person to be tested: match the biometric fusion features of the person to be tested with the biometric features in the identity database to identify the identity information of the person to be tested.

[0053] The specific steps of video detection are as follows: collect the RGB video stream captured by the surveillance camera, and use the pedestrian detection model to infer each image frame of the video stream to obtain the confidence score of the existence of pedestrians and the confidence score of the existence of faces in the image frame. The pedestrian detection model is any one of Mask R-CNN, YOLOv5s, and EfficientDet. The above-mentioned pedestrians refer to people who are moving. The confidence score of pedestrians is CS_person, and the confidence score of faces is CS_face. Considering the pixel quality of the surveillance video and environmental interference, the confidence score threshold of the existence of pedestrians (CT_person) and the confidence score threshold of the existence of faces (CT_face) are defined. If the confidence score of pedestrians and the confidence score of faces in the model inference results exceed the corresponding thresholds, it means that pedestrians or faces are detected. If no pedestrians or faces are detected in the surveillance video, the process ends.

[0054] The specific steps of gait feature extraction are as follows: if a pedestrian is detected (i.e. CS_person>CT_person), the gait biometric features of the pedestrian are extracted. Figure 3As shown in the figure, the effective image area where pedestrians exist is cropped to improve the coverage of the gait feature area. Based on the pedestrian detection frame inferred by the pedestrian detection model, the pedestrian image area is cropped with a given boundary filling pixel size, and the pedestrian image sequence frame group is constructed under the continuous monitoring video stream. The cropped pedestrian image area is image aligned to ensure that the pedestrian feature distribution of the image sequence is similar and the image pixel size is consistent. Image alignment uses means such as feature matching and pixel interpolation. Using behavior recognition models (such as I3D, X3D, slowfast, etc.), the pedestrian image sequence is judged to obtain the action state of the pedestrian in each frame image (such as sitting, standing, lying, walking, etc.), and the pedestrian image sequence in the "walking" action state is selected. To ensure that the image sequence covers the complete gait performance and characteristics, the video length occupied by the image sequence should be no less than 4 seconds. This sequence will be used as the input sample for subsequent pedestrian gait recognition. Models such as GaitSet and GaitCycle are used to obtain the temporal features, spatial features, and deep image features of pedestrian gait. Temporal features include gait cycle, phase information, etc. The gait cycle refers to the time required for a complete step of walking, including support time and swing time, reflecting the rhythmicity of gait; phase information refers to the duration and changes of each stage of the gait cycle, reflecting the specificity of gait. Spatial features include stride length, step width, etc. Deep image features refer to high-dimensional, implicit, deep feature vectors extracted based on convolutional neural networks.

[0055] The facial feature extraction steps are as follows: Figure 4 As shown in the figure, the effective image area with face is cropped to improve the coverage of facial feature area. Based on the face detection frame inferred by the pedestrian detection model, the face image area is cropped with a given boundary filling pixel size. The face image is enhanced by using pixel interpolation method, brightness, contrast and saturation adjustment method to solve the problems of unclear face image or external environment interference and improve image quality. The Zero123++ framework is used to realize the diffusion of face images from a single perspective to multiple perspectives, complete the data fusion and enhancement of multi-angle face information, and effectively improve the significance and specificity of facial features. Finally, FaceNet, DeepFace and other models are used to extract facial features and derive facial feature vectors.

[0056] The specific steps of feature fusion are as follows: some surveillance videos have clearer face images but less gait information; some surveillance videos have complete gait information but no face information (such as walking with the back to the camera); some surveillance videos have both gait information and face information. Formulating a scientific and effective feature fusion strategy can significantly improve the accuracy and applicability of elderly identity recognition. Considering the prominence of gait features and facial features, a weighted fusion strategy is formulated. The prominence of features can be calculated based on the following indicators, including but not limited to: feature similarity, model confidence, image quality score, feature area coverage, etc. The calculation formula is as follows, gait_prominence and face_prominence represent the prominence of gait features and facial features respectively, and [x1,x2,x3,x4], [y1,y2,y3,y4] are parameter influencing factors:

[0057] face_prominence=x1*face_similarity+x2*face_confidence+x3*face_imageQuality+x4*face_featureAreaCoverage

[0058] Among them, face_prominence is the prominence of facial features, face_similarity is the similarity of facial features, face_confidence is the confidence of facial model, face_imageQuality is the facial image quality score, face_featureAreaCoverage is the facial feature area coverage, and [x1,x2,x3,x4] are parameter influencing factors;

[0059] The calculation formula of gait feature significance is:

[0060] gait_prominence=y1*gait_similarity+y2*gait_confidence+y3*gait_imageQuality+y4*gait_featureAreaCoverage

[0061] Among them, gait_prominence is the prominence of gait features, gait_similarity is the similarity of gait features, gait_confidence is the confidence of gait model, gait_imageQuality is the gait image quality score, gait_featureAreaCoverage is the gait feature area coverage, and [y1,y2,y3,y4] are parameter influencing factors.

[0062] According to the specific values ​​of gait_prominence and face_prominence, the feature fusion weighting method is set to reflect the importance of gait features and facial features in identity recognition reasoning. Weighting includes but is not limited to the following methods: linear weighted average, geometric weighting, exponential weighting, nonlinear function weighting, etc. Taking linear weighting as an example, the feature fusion vector feature_fusion calculation formula is:

[0063] feature_fusion=w1*gait_prominence+w2*face_prominence,

[0064] Weight w1=gait_prominence / (gait_prominence+face_prominence)

[0065] Weight w2=face_prominence / (gait_prominence+face_prominence).

[0066] The specific steps to establish a human biometric identity database are:

[0067] The biometric identity database is the unique and only identity authentication credential for each elderly person as the feature matching benchmark for the person to be detected. First, the identity information of all the elderly in the park is counted, including but not limited to: ID number, age, contact information, height, weight, health information, etc. Then the biometric information of the elderly is collected, which is divided into two parts: personal facial image collection and gait video collection during normal walking. The collected biometric information must be of high quality to meet the requirements of subsequent identity feature extraction. High-quality biometric information can improve the performance of the identity recognition fusion model. When collecting facial images, standard collection processes and conditions must be designed to ensure the quality of facial images, such as meeting the following conditions: good lighting conditions, multi-angle face collection, and no decorations. When collecting gait information, it is necessary to ensure that the walking time of the elderly meets specific requirements and covers multiple complete gait cycles; in addition, gait videos from multiple perspectives can be collected, such as gait videos in the sagittal and coronal planes, so as to improve the significance, difference, and robustness of gait information. Then, we extract the biometric features of the elderly, extract facial features based on models such as FaceNet and DeepFace, and obtain gait timing features based on models such as GaitSet and GaitCycle. Finally, we complete the construction of the biometric database of the elderly in the park, such as Figure 5 As shown, each instance of the identity database is the biometric identity of each elderly person, and the identity instance includes the elderly person's identity information, gait feature vector, and facial feature vector.

[0068] The location positioning steps are described by taking the nursing home as an example, as follows:

[0069] The nursing home is divided into multiple physical subspaces within the park area, each subspace is equipped with a surveillance camera, and the video-camera-subspace association relationship is established according to the regional location coding strategy. Based on the reasoning results of the "identity recognition unit", the identity of the elderly in the surveillance video can be recognized, and the association relationship between the elderly identity and the park subspace can be further established, thereby realizing the indoor and outdoor positioning of the elderly.

[0070] Based on the identity recognition model that integrates multi-modal biometrics, this positioning strategy can be implemented with the existing monitoring system to solve many problems existing in traditional indoor and outdoor positioning: 1. The object to be positioned needs to be equipped with or carry specific hardware modules (such as GPS, Bluetooth, WiFi modules, etc.), which cannot be completely imperceptible. The elderly, a special group, have poor compliance with wearable devices; 2. Traditional positioning methods have poor indoor positioning accuracy and it is difficult to determine the location information in the vertical dimension; if you want to achieve higher indoor positioning accuracy, you need to equip Bluetooth / wifi / Lora base stations and corresponding sub-nodes or auxiliary sensor equipment (such as barometers, etc.) over a large area and with high density, which has high planning costs, installation costs and economic costs.

[0071] like Figure 6 As shown, the park is divided into detailed areas according to the activity range or danger level of the elderly, including but not limited to: activity areas, fenced areas, work restricted areas, dangerous areas, etc.

[0072] Coding strategy for indoor and outdoor locations of regions. Any physical subspace in the park, whether indoors or outdoors, can be coded according to Table 1. The first-level code indicates the security level of this space, including but not limited to the following 4 levels (1-safety, 2-low warning, 3-medium warning, 4-high warning); the second-level code represents the specific area to which the space belongs, including but not limited to the following 4 categories (1-activity area, 2-fenced area, 3-work restricted area, 4-dangerous area); the third-level coding sub-area name further refines the area to which the space belongs, such as: east exit, west exit, calligraphy and painting room, gym, pond water area, power distribution room, etc.; the fourth-level code is used to describe the detailed location of this space. Indoor space can be described by building, floor, and room, and outdoor space can be described by the west side of XX building and the east side of XX road; the equipment code indicates the surveillance camera equipped in this space.

[0073] Table 1

[0074]

[0075] An example is as follows: the location code of the space where the camera devId_1 is located is [first-level code: 1-safety; second-level code: 1-activity area; third-level code: calligraphy and painting room; fourth-level code: Building A / 1st floor / 102; device code: devId_1]. If the elderly person old_people_1 is recognized from the camera surveillance video, it means that the detailed location of this elderly person is: activity area, calligraphy and painting room, Building A / 1st floor / 102, and the current area he is in is relatively safe.

[0076] The specific steps of behavioral analysis are:

[0077] like Figure 7As shown in the figure, continuous surveillance video frames are extracted as data input for the elderly behavior analysis model. The behavior analysis model pipeline is built, which consists of two parts. The instance segmentation model (such as Mask R-CNN, YOLOv5+SAM, U-Net, etc.) is used to infer the human body boundary, the external environment boundary, the risk factor boundary, the number of elderly pedestrians and other information; the video action understanding model is used to identify the action state of the elderly (such as walking, sitting, standing, getting up, waving arms, lying, climbing stairs, etc.), the duration of the current action, etc. The above reasoning results, combined with the indoor and outdoor positioning information of the elderly, can be used to detect security events. Position information is very important for behavioral safety analysis, including: a. Reducing the false alarm rate of detection. If some surveillance videos have fewer elderly behavior results and more external environment results, it is easy to generate obvious environmental noise interference. At this time, by combining positioning data, it can be more accurately judged whether the elderly are in the normal activity range, thereby reducing unnecessary alarms; b. Improving detection accuracy. Position information helps the system better understand the behavior patterns of the elderly, thereby improving detection accuracy; c. Facilitating rapid intervention. When safety and abnormal events are detected, accurate positioning makes it easier for managers / caregivers to rush to the scene as soon as possible; d. It is convenient to flexibly define abnormal events. Some institutions or nursing homes have special requirements for safety event detection at specific locations. Positioning information makes the system's definition of abnormal events more flexible and more oriented.

[0078] By combining the above information, i.e., the elderly's location information, external environment boundary information, the elderly's action status, action duration, the number of elderly people, etc., comprehensive and accurate detection of elderly behavioral safety events can be achieved, including but not limited to the following examples:

[0079] (1) If the elderly person's external environment is "ground", the elderly person's action state is "lying" or "struggling", and the action duration is greater than 10 seconds, it is judged as a fall event;

[0080] (2) The location information of the elderly is "pond water area", and the external environment detects "water splashes". The boundary box of the elderly overlaps with the boundary box of the water splashes, and it is determined that the elderly fell into the water;

[0081] (3) If the number of elderly people detected is greater than 6, the overlapping area of ​​the boundary boxes of different elderly people is large, and the duration is greater than 5 minutes, it is determined to be an elderly gathering event.

[0082] The safety detection method of the present invention can significantly improve the accuracy and high applicability of elderly identification through video detection, gait feature extraction, facial feature extraction, feature fusion, establishment of a human biometric identity database, and identification of identity information of the person to be detected. The gait features and facial features of the detected pedestrians are extracted and integrated at the same time, and it is applicable to the following monitoring situations, such as: 1. The elderly are close to the monitoring camera, and only face information exists, but no gait information; 2. The elderly are far away from the monitoring camera, and no face information exists, but only gait information exists; 3. The elderly are at a moderate distance from the monitoring camera, and both face information and gait information exist, so there are many applicable scenarios. The indoor and outdoor position coding strategy is adopted to realize the senseless positioning of the elderly in the park. Solve many problems of traditional indoor and outdoor positioning: 1. The positioned object needs to be equipped with or carry specific hardware modules (such as GPS, Bluetooth, wifi modules, etc.), which cannot be completely senseless. The elderly, a special group, have poor compliance with wearable devices; 2. Traditional positioning methods have poor indoor positioning accuracy and it is difficult to determine the location information in the vertical dimension; if you want to achieve high indoor positioning accuracy, you need to equip Bluetooth / wifi / Lora base stations and corresponding sub-nodes or auxiliary sensor equipment (such as barometers, etc.) over a large area and with high density, which has high planning costs, installation costs and economic costs. By inferring the elderly's positioning information, external environment boundary information, the elderly's action status, action duration, the number of elderly people and other information, it can realize the detection of safety events (falls, falling into water, etc.) and abnormal events (gathering, approaching restricted areas / dangerous areas, etc.), reducing the safety risks of the elderly.

[0083] The above embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which are equivalent modifications and improvements made to the above embodiments based on the essential technology of the present invention, and all of them belong to the protection scope of the present invention.

Claims

1. A safety detection method, characterized in that: The following steps are involved: Identity recognition: extract gait and facial multimodal feature vectors through surveillance video, perform multimodal biometric recognition, match the recognition results with the human biometric database, and realize human identity recognition; Position positioning: Divide the area where the person is located into multiple physical subspaces, each subspace is equipped with a surveillance camera, and establish a video-camera-subspace association relationship according to the regional location coding strategy, and obtain the location information of the person with the corresponding identity based on the identity recognition result; Behavior analysis: Using surveillance video as input data for human behavior analysis, an instance segmentation model is used to obtain information on human body boundaries, external environment boundaries, risk factor boundaries, and the number of elderly pedestrians. A video action understanding model is used to identify human action states and action durations. Human behavior events are judged based on the location information, action states, action durations, external environment boundaries, risk factor boundaries, and the number of elderly pedestrians. Abnormal event and dangerous event alarm: When a person's behavior is dangerous or abnormal, an alarm will be issued and the person's identity information and corresponding location information will be output.

2. The safety detection method according to claim 1, characterized in that: In the position positioning step, the area where the person is located is divided into multiple physical subspaces, specifically: the area where the person is located is divided into at least two of the following: an activity area, a fenced area, a work restricted area, a dangerous area, a water area, and a staircase.

3. The safety detection method according to claim 1, characterized in that: In the position positioning step, according to the regional position coding strategy, the video-camera-subspace association relationship is established as follows: multiple cameras are encoded separately, and each camera is multi-level encoded, and the multi-level encoding includes at least two of the security level, area name, sub-area name, and detailed location.

4. The safety detection method according to claim 1, characterized in that: In the behavior analysis step, the video action understanding model is used to identify the action states of a person, and the action states include walking, sitting, standing, getting up, waving arms, squatting, climbing stairs, and lying down.

5. The safety detection method according to claim 1, characterized in that: In the behavior analysis step, the dangerous and abnormal situations of the behavior events include falling, falling into water, gathering and approaching dangerous areas.

6. The safety detection method according to claim 1, characterized in that: The identity identification step specifically includes Video detection: Obtain camera surveillance video, use the pedestrian detection model to infer each image frame of the surveillance video stream, and obtain the confidence score of the existence of pedestrians and the confidence score of the existence of faces in the image frame. When the confidence score of pedestrians and / or the confidence score of faces is greater than the threshold, it is determined that pedestrians and / or faces are detected; Gait feature extraction: When a pedestrian is detected, the effective image area where the pedestrian exists is cropped, and a pedestrian image sequence frame group is constructed. The cropped pedestrian image area is aligned, and the pedestrian image sequence is judged using the behavior recognition model to obtain the action state of the pedestrian in each frame image. The pedestrian image sequence in the "walking" action state is selected, and the pedestrian gait features are obtained based on the pedestrian image sequence in the "walking" action state; Facial feature extraction: When a face is detected, the effective image area where the face exists is cropped, the face image is enhanced, the face image is diffused in multiple dimensions, the data fusion of multi-angle face information is completed, and the facial features of the fused image are extracted; Feature fusion: Calculate the gait feature significance and the facial feature significance, determine the feature fusion weighting method according to the numerical values ​​of the gait feature significance and the facial feature significance, and fuse the gait features and facial features according to the determined feature fusion weighting method to obtain a feature fusion vector; Establishing a human biometric identity database: Collecting each person's identity information and biological information to establish a human biometric identity database, the biological information includes facial images and gait information; Identify the identity information of the person to be tested: match the biometric fusion features of the person to be tested with the biometric features in the identity database to identify the identity information of the person to be tested.

7. The safety detection method according to claim 6, characterized in that: In the gait feature extraction step, the construction of the pedestrian image sequence frame group is specifically: using a pedestrian detection frame based on pedestrian detection model reasoning, giving a boundary filling pixel size, cropping the pedestrian image area, and constructing a pedestrian image sequence frame group under a continuous monitoring video stream.

8. The safety detection method according to claim 6, characterized in that: In the gait feature extraction step, the image alignment operation performed on the cropped pedestrian image region is specifically performed by using feature matching and pixel interpolation.

9. The safety detection method according to claim 6, characterized in that: In the feature fusion step, the gait feature prominence and the facial feature prominence are calculated according to the corresponding feature approximation, model confidence, image quality score and feature area coverage.

10. A safety detection system for implementing the safety detection method according to any one of claims 1 to 6, characterized in that: The safety detection system comprises The acquisition module, multiple cameras located in different areas collect surveillance videos of the corresponding areas; An analysis module, which analyzes the identity information of the person in the surveillance video, obtains the location information of the person according to the coding of the camera, and performs behavior analysis according to the location information and the surveillance video; The output module will sound an alarm when a person's behavior is dangerous or abnormal, and output the person's identity information and corresponding location information.

Images