Computer vision-based intelligent monitoring system for campus safety

Abstract

The present application belongs to the field of intelligent monitoring, and specifically is a campus safety intelligent monitoring system based on computer vision, a video monitoring module, an image enhancement module, an intelligent face recognition module, a behavior recognition module based on skeleton, an alarm module and a central control module; the present application adopts a face recognition model based on multi-angle panoramic three-dimensional optimization, integrates feature information of different angles and depths, reduces the degree of influence by factors such as light and posture, and through feature simplification, scaling and selection, removes redundant information, and more accurately performs face recognition; the present application adopts a multi-modal feature extraction and fusion method based on channel convolution attention module optimization, extracts spatial position and human key points at high resolution, introduces Gaussian distribution to construct a skeleton heat map, and adopts an action recognition model based on channel attention optimization and an RGB feature extraction model to perform multi-modal feature extraction and fusion on pedestrian images, and realizes high-accuracy and high-efficiency behavior recognition.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligent monitoring, and particularly relates to a campus safety intelligent monitoring system based on computer vision. BACKGROUND

[0002] The background of the campus safety monitoring system is derived from the importance and demand for campus safety issues. With the development of society and the progress of technology, the campus safety issue is increasingly concerned. School administrators need to ensure the safety of teachers, students and staff in the campus, and also need to maintain the order of the campus and manage various activities in the campus, so higher requirements are put forward for the campus safety management. However, the existing monitoring has low accuracy of video face recognition, and the recognition of dangerous behaviors in the campus relies on single modal action recognition, which has high error rate and low efficiency. SUMMARY

[0003] In view of the above situation, in order to overcome the defects of the prior art, in view of the low accuracy of video face recognition of the existing monitoring, the application creatively adopts a face recognition model based on multi-angle panoramic stereo optimization. By integrating feature information of different angles and depths, the model obtains more comprehensive and accurate face representation, reduces the degree of influence of factors such as light and posture, and removes redundant information through feature simplification, scaling and selection, so as to more accurately identify, reduce false positives and false alarms, and improve the reliability of the system. In view of the problem that the existing monitoring relies on single modal action recognition for the recognition of dangerous behaviors in the campus, and has high error rate and low efficiency, the application adopts a multi-modal feature extraction and fusion method based on channel convolution attention module optimization. A high-resolution network is used to extract spatial position information of the image, a top-down pose estimation algorithm Lite-HRNet is used to identify human key points, a Gaussian distribution is introduced to construct a skeleton heat map, and a motion recognition model based on channel attention optimization is used to extract skeleton modal features of the skeleton heat map. An RGB feature extraction model based on channel attention optimization is used to extract RGB features of the pedestrian image. Finally, the motion is recognized by the fused features, so as to realize high-accuracy and high-efficiency behavior recognition.

[0004] The technical scheme provided by the application is as follows: a campus safety intelligent monitoring system based on computer vision, comprising a video monitoring module, an image enhancement module, an intelligent face recognition module, a behavior recognition module based on a skeleton, an alarm module and a central control module.

[0005] The video monitoring module collects pedestrian images in the campus through a high-definition camera, and transmits the pedestrian images to the image enhancement module.

[0006] The image enhancement module performs spatial domain-based enhancement and frequency domain-based enhancement processing on the pedestrian image to obtain an enhanced image, and sends the enhanced image to the intelligent face recognition module and the skeleton-based behavior recognition module.

[0007] The spatial domain-based enhancements include gamma correction and histogram equalization;

[0008] The frequency-domain-based enhancement includes Gaussian low-pass filtering;

[0009] The intelligent face recognition module uses a face recognition model based on multi-angle panoramic stereo optimization to intelligently identify pedestrians in enhanced images.

[0010] The skeleton-based behavior recognition module performs behavior recognition on pedestrians in the enhanced image, including a human target detection unit, a key point recognition unit, and an action recognition unit.

[0011] The human target detection unit detects human targets in the enhanced image using the YOLOX target detection algorithm;

[0012] The key point recognition unit performs human key point recognition on human targets using the human key point recognition method of the lightweight high-resolution network Lite-HRNet.

[0013] The action recognition unit identifies pedestrian behavior and includes a skeletal modality feature extraction unit, an RGB modality feature extraction unit, and a feature fusion unit.

[0014] The skeletal modality feature extraction unit extracts skeletal modality features using a Gaussian distribution-based channel attention optimization feature extraction method.

[0015] The RGB modal feature extraction unit extracts RGB modal features using a channel attention-optimized feature extraction method.

[0016] The feature fusion unit combines skeletal modal features and RGB modal features, activates and outputs them to obtain pedestrian behavior recognition results.

[0017] The alarm module classifies the pedestrian behavior recognition results and issues an alarm when a dangerous behavior is identified.

[0018] The central control module controls and stores information for all modules.

[0019] Furthermore, in the intelligent face recognition module, the data processing of the face recognition model based on multi-angle panoramic stereoscopic optimization specifically includes the following steps:

[0020] Step S1: Multi-angle convolution. The face images in the enhanced image are acquired through a angles, and a identical graph convolutional networks are used to convolve the face images from a different angles to obtain a initial feature maps with dimensions H×W×C, where the face images from a angles fully cover the face.

[0021] Step S2: Feature concatenation. Concatenate the a initial feature maps while keeping the height and width unchanged. The number of channels becomes a times the original, i.e., aC, to obtain the full feature map.

[0022] Step S3: Redundant information removal: Perform a 1×1 point convolution on the full feature map to obtain a simplified feature map.

[0023] ;

[0024] ;

[0025] In the formula, For simplified feature maps, For the image The coordinates in the nth channel are elements, The weights of the convolution kernel for the nth channel of the reduced feature map in the kth channel of the full feature map are: The coordinates of the k-th channel of the full feature map are: Element;

[0026] Step S4: Feature downsampling. Global depthwise convolution is used to downsample the simplified feature map, resulting in a simplified feature vector.

[0027] ;

[0028] ;

[0029] In the formula, This is the value of the b-th channel after performing a global depthwise convolution on the simplified feature map, representing the information of the b-th channel. The depthwise convolution kernel parameter matrix represents the b-th channel. This represents the feature of the b-th channel in the simplified feature map. This represents the simplified feature vector obtained from global depthwise convolution;

[0030] Step S5: Feature vector compression. Compress the simplified feature vector to obtain the compressed feature vector:

[0031] ;

[0032] In the formula, Represents the activation function. The parameter matrix representing the fully connected layer. Represents a compressed feature vector;

[0033] Step S6: Weight vector calculation:

[0034] ;

[0035] In the formula, The parameter matrix representing the fully connected layer. Let C be the weight vector of the b-th dimension, where m ∈ [1,a] , For activation functions;

[0036] Step S7: Feature selection and fusion to obtain facial features:

[0037] ;

[0038] ;

[0039] In the formula, Representing the An initial feature map, Let C be the weight vector of the b-th dimension. Facial features that filter out redundant information;

[0040] Step S8: Face recognition, which obtains the face recognition result by matching facial features with the pedestrian's identity.

[0041] Furthermore, in the skeletal modality feature extraction unit, the channel attention optimization feature extraction method based on Gaussian distribution specifically includes the following steps:

[0042] Step Q1: Constructing a keypoint heatmap based on Gaussian distribution:

[0043] heatmap (x,y) = e - D((x,y),seg[ a k , b k ] ) 2 2* σ 2 *min( ca k ,c b k ) ;

[0044] In the formula, It is the standard deviation of the Gaussian distribution. For the coordinates of the key points, To start from the key point to line segment seg[ a k , b k ] distance, Key point and The confidence score, Represents a heatmap of key points;

[0045] Step Q2: 3D heatmap construction, which is achieved by stacking key point heatmaps along the time dimension to obtain a 3D heatmap;

[0046] Step Q3: Model building, among which The model consists of multiple ResNet layers, and each ResNet layer includes multiple bottleneck blocks;

[0047] Step Q4: Model optimization, using channel attention mechanism The bottleneck block of the ResNet layer in the model is replaced, and the channel attention mechanism is handled as follows:

[0048] ;

[0049] M S (F)= M C (F)⊗ σ ( f 7 × 7 ([MaxPool( M C (F));AvgPool( M C (F))])) ;

[0050] In the formula, Represents the input to the ResNet layer. Represents multi-level vector machine processing, Represents max pooling. Represents average pooling. Represents element-wise multiplication. Represents the Sigmoid activation function. As an intermediate variable, Represents a 7×7 convolution. The calculation results represent the channel attention mechanism;

[0051] Step Q5: Feature extraction, through optimization The model extracts and activates features from 3D heatmaps to obtain skeletal modal features;

[0052] Furthermore, in the RGB modal feature extraction unit, the channel attention optimization feature extraction method specifically includes the following steps:

[0053] Step P1: Slowonly model construction, where the Slowonly model includes multiple ResNet layers, and each ResNet layer includes multiple bottleneck blocks;

[0054] Step P2: Slowonly model optimization, using channel attention mechanism to replace the bottleneck block of the ResNet layer in the Slowonly model;

[0055] Step P3: Feature extraction. Using an optimized Slowonly model, feature extraction and activation are performed on the enhanced image to obtain RGB modal features.

[0056] The beneficial effects achieved by the present invention using the above solution are as follows:

[0057] (1) In view of the problem that the accuracy of video face recognition is low in existing monitoring, the present invention creatively adopts a face recognition model based on multi-angle panoramic stereo optimization. By integrating feature information from different angles and depths, the model obtains a more comprehensive and accurate face representation, reducing the degree of influence from factors such as lighting and posture. At the same time, by simplifying, scaling and selecting features, redundant information is removed, which can make the recognition more accurate, reduce false alarms and false alarms, and improve the reliability of the system.

[0058] (2) In view of the existing monitoring, the identification of dangerous behaviors on campus relies on single-modality action recognition, which has a high error rate and low efficiency. This invention adopts a multimodal feature extraction and fusion method based on channel convolutional attention module optimization. It uses a high-resolution network to extract spatial location information from the image, uses the top-down pose estimation algorithm Lite-HRNet to identify human key points, introduces Gaussian distribution to construct skeletal heatmaps, uses a channel attention-optimized action recognition model to extract skeletal modal features from the skeletal heatmaps, and uses a channel attention-optimized RGB feature extraction model to extract RGB features from pedestrian images. Finally, the fused features are used to identify the actions, achieving high accuracy and high efficiency in behavior recognition. Attached Figure Description

[0059] Figure 1 A block diagram of the computer vision-based intelligent campus security monitoring system provided by the present invention;

[0060] Figure 2 This is a schematic diagram of the data processing flow of a face recognition model based on multi-angle panoramic stereo optimization.

[0061] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0062] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0063] Example 1, see Figure 1 The intelligent campus security monitoring system based on computer vision provided by this invention is characterized by:

[0064] It includes a video surveillance module, an image enhancement module, an intelligent face recognition module, a skeleton-based behavior recognition module, an alarm module, and a central control module;

[0065] The video surveillance module captures pedestrian images on campus using a high-definition camera and transmits the pedestrian images to the image enhancement module.

[0066] The image enhancement module performs spatial domain-based enhancement and frequency domain-based enhancement processing on the pedestrian image to obtain an enhanced image, and sends the enhanced image to the intelligent face recognition module and the skeleton-based behavior recognition module.

[0067] The spatial domain-based enhancements include gamma correction and histogram equalization;

[0068] The frequency-domain-based enhancement includes Gaussian low-pass filtering;

[0069] The intelligent face recognition module uses a face recognition model based on multi-angle panoramic stereo optimization to intelligently identify pedestrians in enhanced images and obtain pedestrian identification results.

[0070] The skeleton-based behavior recognition module performs behavior recognition on pedestrians in the enhanced image, including a human target detection unit, a key point recognition unit, and an action recognition unit.

[0071] The alarm module classifies the pedestrian behavior recognition results and issues an alarm when a dangerous behavior is identified.

[0072] The central control module controls and stores information for all modules.

[0073] Example 2: This example is based on the above examples. In the skeleton-based behavior recognition module,

[0074] The human target detection unit detects human targets in the enhanced image using the YOLOX target detection algorithm;

[0075] The key point recognition unit performs human key point recognition on human targets using the human key point recognition method of the lightweight high-resolution network Lite-HRNet.

[0076] The action recognition unit identifies pedestrian behavior and includes a skeletal modality feature extraction unit, an RGB modality feature extraction unit, and a feature fusion unit.

[0077] The skeletal modality feature extraction unit extracts skeletal modality features using a Gaussian distribution-based channel attention optimization feature extraction method.

[0078] The RGB modal feature extraction unit extracts RGB modal features using a channel attention-optimized feature extraction method.

[0079] The feature fusion unit combines skeletal modal features and RGB modal features, activates and outputs them to obtain pedestrian behavior recognition results.

[0080] Example 3, see Figure 2 Based on the above embodiments, this embodiment, in the intelligent face recognition module, specifically includes the following steps for data processing of a face recognition model based on multi-angle panoramic stereoscopic optimization:

[0081] Step S1: Multi-angle convolution. The face images in the enhanced image are acquired through a angles, and a identical graph convolutional networks are used to convolve the face images from a different angles to obtain a initial feature maps with dimensions H×W×C, where the face images from a angles fully cover the face.

[0082] Step S2: Feature concatenation. Concatenate the a initial feature maps while keeping the height and width unchanged. The number of channels becomes a times the original, i.e., aC, to obtain the full feature map.

[0083] Step S3: Redundant information removal: Perform a 1×1 point convolution on the full feature map to obtain a simplified feature map.

[0084] ;

[0085] ;

[0086] In the formula, For simplified feature maps, For the image The coordinates in the nth channel are elements, The weights of the convolution kernel for the nth channel of the reduced feature map in the kth channel of the full feature map are: The coordinates of the k-th channel of the full feature map are: Element;

[0087] Step S4: Feature downsampling. Global depthwise convolution is used to downsample the simplified feature map, resulting in a simplified feature vector.

[0088] ;

[0089] ;

[0090] In the formula, This is the value of the b-th channel after performing a global depthwise convolution on the simplified feature map, representing the information of the b-th channel. The depthwise convolution kernel parameter matrix represents the b-th channel. This represents the feature of the b-th channel in the simplified feature map. This represents the simplified feature vector obtained from global depthwise convolution;

[0091] Step S5: Feature vector compression. Compress the simplified feature vector to obtain the compressed feature vector:

[0092] ;

[0093] In the formula, Represents the activation function. The parameter matrix representing the fully connected layer. Represents a compressed feature vector;

[0094] Step S6: Weight vector calculation:

[0095] ;

[0096] In the formula, The parameter matrix representing the fully connected layer. Let C be the weight vector of the b-th dimension, where m ∈ [1,a] , For activation functions;

[0097] Step S7: Feature selection and fusion to obtain facial features:

[0098] ;

[0099] ;

[0100] In the formula, Representing the An initial feature map, Let C be the weight vector of the b-th dimension. Facial features that filter out redundant information;

[0101] Step S8: Face recognition, which obtains the pedestrian identity recognition result by matching facial features with the pedestrian's identity.

[0102] By performing the above operations, this invention creatively adopts a face recognition model based on multi-angle panoramic stereo optimization to address the problem of low accuracy in existing surveillance video face recognition. By integrating feature information from different angles and depths, the model obtains a more comprehensive and accurate face representation, reducing the degree of influence from factors such as lighting and posture. At the same time, through feature simplification, scaling, and selection, redundant information is removed, enabling more accurate recognition, reducing false alarms and false alarms, and improving the reliability of the system.

[0103] Example 4: This example is based on the above examples.

[0104] In the skeletal modality feature extraction unit, the channel attention optimization feature extraction method based on Gaussian distribution specifically includes the following steps:

[0105] Step Q1: Constructing a keypoint heatmap based on Gaussian distribution:

[0106] heatmap (x,y) = e - D((x,y),seg[ a k , b k ] ) 2 2* σ 2 *min( ca k ,c b k ) ;

[0107] In the formula, It is the standard deviation of the Gaussian distribution. For the coordinates of the key points, To start from the key point to line segment seg[ a k , b k ] distance, Key point and The confidence score, Represents a heatmap of key points;

[0108] Step Q2: 3D heatmap construction, which is achieved by stacking key point heatmaps along the time dimension to obtain a 3D heatmap;

[0109] Step Q3: Model building, among which The model consists of multiple ResNet layers, and each ResNet layer includes multiple bottleneck blocks;

[0110] Step Q4: Model optimization, using channel attention mechanism The bottleneck block of the ResNet layer in the model is replaced, and the channel attention mechanism is handled as follows:

[0111] ;

[0112] M S (F)= M C (F)⊗ σ ( f 7 × 7 ([MaxPool( M C (F));AvgPool( M C (F))])) ;

[0113] In the formula, Represents the input to the ResNet layer. Represents multi-level vector machine processing, Represents max pooling. Represents average pooling. Represents element-wise multiplication. Represents the Sigmoid activation function. As an intermediate variable, Represents a 7×7 convolution. The calculation results represent the channel attention mechanism;

[0114] Step Q5: Feature extraction, through optimization The model extracts and activates features from 3D heatmaps to obtain skeletal modal features.

[0115] Example 5: Based on the above examples, in the RGB modal feature extraction unit, the channel attention optimization feature extraction method specifically includes the following steps:

[0116] Step P1: Slowonly model construction, where the Slowonly model includes multiple ResNet layers, and each ResNet layer includes multiple bottleneck blocks;

[0117] Step P2: Slowonly model optimization, using channel attention mechanism to replace the bottleneck block of the ResNet layer in the Slowonly model;

[0118] Step P3: Feature extraction. Using an optimized Slowonly model, feature extraction and activation are performed on the enhanced image to obtain RGB modal features.

[0119] By performing the above operations, this invention addresses the problem that existing monitoring relies on single-modal action recognition for identifying dangerous behaviors on campus, resulting in a high error rate and low efficiency. It employs a multimodal feature extraction and fusion method based on channel convolutional attention module optimization. This method uses a high-resolution network to extract spatial location information from images, a top-down pose estimation algorithm (Lite-HRNet) to identify key points on the human body, introduces a Gaussian distribution to construct a skeletal heatmap, and uses a channel attention-optimized action recognition model to extract skeletal modal features from the heatmap. It also uses a channel attention-optimized RGB feature extraction model to extract RGB features from pedestrian images. Finally, the fused features are used to identify actions, achieving highly accurate and efficient behavior recognition.

[0120] Example 6: Based on the above examples, the pedestrian behavior recognition results and classification details are as follows:

[0121] Category of dangerous behavior:

[0122] Violent acts: such as fighting, pushing, and injuring others;

[0123] Illegal climbing: scaling over school walls and entering restricted areas;

[0124] Unusual gatherings: Students gathering or causing congestion in a certain area for an extended period of time;

[0125] Carrying dangerous items: knives, sticks, etc.;

[0126] Fall / Unconsciousness: such as a student falling or remaining motionless for an extended period of time.

[0127] Example 7: This example is based on the above examples, and describes the actual application scenario of the present invention.

[0128] Entrance and exit monitoring:

[0129] Facial recognition cameras are deployed at the school gates and dormitory entrances to record all people entering and exiting;

[0130] If an unauthorized person attempts to enter, an audible and visual alarm will be triggered, and the access control system will be prevented from opening.

[0131] Teaching area surveillance:

[0132] Deploy cameras to detect dangerous behavior in high-traffic areas such as corridors and stairwells of teaching buildings;

[0133] When the system detects abnormal gatherings, pushing, or other disruptive behaviors, it issues an early warning to prevent stampedes.

[0134] Monitoring of playground and public areas:

[0135] 360° panoramic cameras are deployed on the playground to monitor a large area of ​​activity.

[0136] When violent behavior or a fall is detected, the system automatically locates the incident and notifies security personnel.

[0137] Restricted area monitoring:

[0138] In restricted areas such as laboratories and computer rooms, the system combines access control and facial recognition to prevent unauthorized entry;

[0139] When unauthorized access is detected, the system will automatically record the incident and issue an alarm.

[0140] Example 8: Based on the above examples, this example optimizes and further expands the functionality of the face recognition system.

[0141] Identity verification process:

[0142] Students and faculty members must have their facial information entered into a database in advance, while visitors must register their identity and have their faces photographed at the entrance.

[0143] The system compares people entering and leaving the campus in real time to determine whether they are authorized personnel;

[0144] Abnormal identity detection:

[0145] Unauthorized personnel intrusion: The system identifies unregistered personnel and triggers an alarm;

[0146] Blacklist comparison: Match against a blacklist database (including long-staying individuals, potential threats, etc.).

[0147] Attendance function:

[0148] Facial recognition is used to automatically record student attendance data, avoiding omissions caused by manual registration.

[0149] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0150] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0151] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A campus security intelligent monitoring system based on computer vision, characterized in that: It includes a video surveillance module, an image enhancement module, an intelligent face recognition module, a skeleton-based behavior recognition module, an alarm module, and a central control module; The video surveillance module captures pedestrian images on campus using a high-definition camera and transmits the pedestrian images to the image enhancement module. The image enhancement module performs spatial domain-based enhancement and frequency domain-based enhancement processing on the pedestrian image to obtain an enhanced image, and sends the enhanced image to the intelligent face recognition module and the skeleton-based behavior recognition module. The spatial domain-based enhancements include gamma correction and histogram equalization; The frequency-domain-based enhancement includes Gaussian low-pass filtering; The intelligent face recognition module uses a face recognition model based on multi-angle panoramic stereo optimization to intelligently identify pedestrians in enhanced images and obtain pedestrian identification results. In the intelligent face recognition module, the data processing of the face recognition model based on multi-angle panoramic stereoscopic optimization specifically includes the following steps: Step S1: Multi-angle convolution. The face images in the enhanced image are acquired through a angles, and a identical graph convolutional networks are used to convolve the face images from a different angles to obtain a initial feature maps with dimensions H×W×C, where the face images from a angles fully cover the face. Step S2: Feature concatenation. Concatenate the a initial feature maps while keeping the height and width unchanged. The number of channels becomes a times the original, i.e., aC, to obtain the full feature map. Step S3: Redundant information removal: Perform a 1×1 point convolution on the full feature map to obtain a simplified feature map. ； ； In the formula, For simplified feature maps, For the image The coordinates in the nth channel are elements, The weights of the convolution kernel for the nth channel of the reduced feature map in the kth channel of the full feature map are: The coordinates of the k-th channel of the full feature map are: Element; Step S4: Feature downsampling. Global depthwise convolution is used to downsample the simplified feature map, resulting in a simplified feature vector. ； ； In the formula, This is the value of the b-th channel after performing a global depthwise convolution on the simplified feature map, representing the information of the b-th channel. The depthwise convolution kernel parameter matrix represents the b-th channel. This represents the feature of the b-th channel in the simplified feature map. This represents the simplified feature vector obtained from global depthwise convolution; Step S5: Feature vector compression. Compress the simplified feature vector to obtain the compressed feature vector: ； In the formula, Represents the activation function. The parameter matrix representing the fully connected layer. Represents a compressed feature vector; Step S6: Calculate the weight vector based on the compressed feature vector: ； In the formula, The parameter matrix representing the fully connected layer. Let C be the weight vector of the b-th dimension, where , For activation functions; Step S7: Feature selection and fusion to obtain facial features: ； ； In the formula, Representing the An initial feature map, Let C be the weight vector of the b-th dimension. Facial features that filter out redundant information; Step S8: Face recognition, which involves matching facial features with pedestrian identities to obtain pedestrian identification results; The skeleton-based behavior recognition module performs behavior recognition on pedestrians in the enhanced image, including a human target detection unit, a key point recognition unit, and an action recognition unit. The alarm module classifies the pedestrian behavior recognition results and issues an alarm when a dangerous behavior is identified. The central control module controls and stores information for all modules.

2. The campus security intelligent monitoring system based on computer vision according to claim 1, characterized in that: The human target detection unit detects human targets in the enhanced image using the YOLOX target detection algorithm; The key point recognition unit performs human key point recognition on human targets using the human key point recognition method of the lightweight high-resolution network Lite-HRNet. The action recognition unit identifies pedestrian behavior and includes a skeletal modality feature extraction unit, an RGB modality feature extraction unit, and a feature fusion unit. The skeletal modality feature extraction unit extracts skeletal modality features using a Gaussian distribution-based channel attention optimization feature extraction method. The RGB modal feature extraction unit extracts RGB modal features using a channel attention-optimized feature extraction method. The feature fusion unit combines skeletal modal features and RGB modal features, activates and outputs them to obtain pedestrian behavior recognition results.

3. The campus security intelligent monitoring system based on computer vision according to claim 2, characterized in that: In the skeletal modality feature extraction unit, the channel attention optimization feature extraction method based on Gaussian distribution specifically includes the following steps: Step Q1: Constructing a keypoint heatmap based on Gaussian distribution: ； In the formula, It is the standard deviation of the Gaussian distribution. For the coordinates of the key points, To start from the key point to line segment distance, Key point and The confidence score, Represents a heatmap of key points; Step Q2: 3D heatmap construction, which is achieved by stacking key point heatmaps along the time dimension to obtain a 3D heatmap; Step Q3: PoseC3D model building, where the PoseC3D model includes multiple ResNet layers, and each ResNet layer includes multiple bottleneck blocks; Step Q4: PoseC3D model optimization, using channel attention mechanism to replace the bottleneck block of the ResNet layer in the PoseC3D model; Step Q5: Feature extraction. Using the optimized PoseC3D model, feature extraction and activation are performed on the 3D heatmap to obtain skeletal modal features.

Images