Low-cost gait recognition and suspicious behavior detection methods

Abstract

This invention relates to a low-cost gait recognition and suspicious behavior detection method, belonging to the field of image processing. The method utilizes existing surveillance cameras and deep learning algorithms to achieve automatic recognition and alerting of pedestrian behavior. Specific steps include: using a camera to capture pedestrian targets within a monitored area in real time, dynamically tracking the targets, and continuously cropping and saving pedestrian images within the anchor frame; performing edge detection processing on the saved images to extract the main contour information; stacking N adjacent consecutive images to form a three-dimensional tensor; inputting the stacked image into a pre-trained classification system based on a convolutional neural network to analyze the behavioral features in the image sequence, determine whether the target behavior meets the suspicious behavior criteria, and further determine the specific behavior category. This invention has advantages such as low cost, high accuracy, and strong real-time performance.

Description

Technical Field

[0001] This invention belongs to the field of image processing and relates to a low-cost method for gait recognition and suspicious behavior detection. Background Technology

[0002] Suspicious behavior detection technology based on deep gait recognition has made significant progress in recent years, and its current development status is mainly reflected in the following aspects:

[0003] First, the introduction of deep learning technology has significantly improved the accuracy and robustness of gait recognition. Through deep learning models such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs), gait features can be extracted and analyzed more effectively, maintaining high recognition accuracy even in complex backgrounds, changing lighting conditions, and occlusion. This allows gait recognition to move beyond traditional simple feature extraction, greatly enhancing the ability to detect suspicious behavior.

[0004] Secondly, the construction of large-scale datasets provides crucial support for the training and evaluation of gait recognition models. Currently, several publicly available gait datasets (such as CASIA-B and OU-ISIR) cover gait data under different environments, angles, and occlusion conditions, providing a solid foundation for algorithm optimization and practical application testing. The continuous expansion and enrichment of these datasets enable models to maintain strong generalization capabilities in various complex scenarios.

[0005] Furthermore, multimodal fusion is becoming a development trend. To improve recognition accuracy and address the challenges of various environments, deep gait recognition is increasingly combining with other biometrics (such as face and posture) and even integrating behavioral patterns from videos to construct a more comprehensive feature representation. This fusion approach has shown great potential in public safety and intelligent surveillance, enabling more accurate identification and analysis of suspicious behavior.

[0006] However, current technologies still face challenges related to data privacy and ethics. The application of gait recognition involves continuous monitoring of individual behavior, which may raise privacy concerns. Researchers are gradually exploring privacy protection solutions to ensure the compliance of data use.

[0007] The lidar-based gait recognition system utilizes lidar sensors to generate high-precision three-dimensional spatial data to analyze human gait characteristics. Its main technical features are as follows:

[0008] 3D point cloud data acquisition: LiDAR generates high-resolution 3D point cloud data by emitting laser pulses and receiving their reflected signals. This data can accurately reflect the human body's position, shape, and dynamic changes in space, thus supporting the acquisition of gait features.

[0009] Point cloud preprocessing: The raw point cloud data generated by LiDAR typically contains noise and redundant information, therefore denoising, filtering, and downsampling are required to improve data quality and computational efficiency. Simultaneously, point cloud segmentation technology is also used to extract point cloud information of the target human body from the background.

[0010] Human posture modeling and tracking: By extracting geometric features from point cloud data, human posture can be modeled and tracked. Methods such as joint detection and skeleton reconstruction are typically used to generate a 3D representation of the human posture. This posture data provides more accurate dynamic features for gait recognition.

[0011] Gait feature extraction: Based on human posture, gait features are extracted using deep learning or geometric feature analysis methods. Gait features include stride length, stride width, and joint angle changes, which can effectively describe an individual's dynamic characteristics. This process often uses convolutional neural networks (CNNs) or graph convolutional networks (GCNs) to learn and model the data.

[0012] Behavior analysis and recognition: This involves classifying and recognizing extracted gait features, and training models to identify specific gait patterns. Algorithms such as Support Vector Machines (SVM) and Recurrent Neural Networks (RNN) are typically used to map different gait features to specific identities or behavioral patterns to achieve gait recognition.

[0013] Multi-sensor fusion: To improve the adaptability and recognition accuracy of the system, LiDAR is often combined with other sensors (such as cameras and inertial sensors) to achieve multi-modal data fusion. Multi-modal data can compensate for the shortcomings of a single sensor and enhance the stability and robustness of the system in complex scenarios.

[0014] Real-time performance and computational optimization: LiDAR generates large amounts of data and involves complex calculations, making real-time performance crucial for gait recognition systems. Employing efficient point cloud processing algorithms, parallel computing techniques (such as GPU acceleration), and optimized network structures can significantly improve the system's real-time response capabilities.

[0015] By implementing the above technical points, the gait recognition system based on LiDAR can accurately capture and analyze human gait characteristics in various scenarios, providing strong technical support for applications such as identity recognition and abnormal behavior monitoring.

[0016] Gait recognition systems based on lidar have certain advantages in detection accuracy and data richness, but they also have some drawbacks and limitations:

[0017] High cost: LiDAR equipment is expensive, and the cost of system construction and maintenance is relatively high, especially when deployed on a large scale. This high cost can become a major obstacle to practical application.

[0018] Equipment size and complexity: LiDAR devices are usually large and complex, suitable for fixed monitoring scenarios, but lack in portability and flexibility, so they are not easy to apply to lightweight and highly mobile scenarios.

[0019] Environmental dependence: LiDAR systems perform well under specific conditions, but are easily interfered with in rainy or snowy weather, strong sunlight, or environments with a lot of suspended particles in the air, which affects the accuracy and stability of the data.

[0020] Large data processing volume: The amount of data collected by LiDAR is huge, especially when capturing complex movements or high-frequency sampling, which requires a lot of computing resources for data processing. This places high demands on system hardware and may cause latency issues in real-time application scenarios.

[0021] Privacy and security concerns: LiDAR technology can capture detailed three-dimensional information about the human body, potentially involving personal privacy data. Although LiDAR does not directly capture image information, the high-precision three-dimensional data it generates may still pose privacy challenges, especially with its widespread use in public places.

[0022] While LiDAR has unique advantages in terms of accuracy and applicability in gait recognition, its shortcomings in terms of cost, equipment complexity, environmental adaptability, and privacy issues need to be further addressed before it can be more widely used in gait recognition and suspicious behavior detection.

[0023] This invention primarily addresses the problems of high equipment cost, large and complex size, and privacy concerns in existing LiDAR-based gait recognition systems. It fully utilizes existing camera-based monitoring systems and combines deep learning and computer vision technologies to achieve similar functionality and performance metrics at a lower cost. Summary of the Invention

[0024] In view of this, the purpose of this invention is to provide a low-cost gait recognition and suspicious behavior detection method, aiming to address the shortcomings of current LiDAR-based gait recognition systems, such as high equipment cost, large size, complex structure, and privacy protection. By fully utilizing existing camera monitoring systems and combining deep learning and computer vision technologies, the method achieves abnormal behavior detection at a lower cost and reaches a performance level similar to existing technologies.

[0025] To achieve the above objectives, the present invention provides the following technical solution:

[0026] A low-cost gait recognition and suspicious behavior detection method includes the following steps:

[0027] Use cameras to capture pedestrian targets in real time within the monitored area;

[0028] Dynamically track pedestrian targets and continuously capture and save pedestrian images within the anchor frame;

[0029] Edge detection is performed on the saved image to extract its contour information;

[0030] Select N consecutive adjacent images and stack them to form a three-dimensional tensor;

[0031] The stacked images are input into a pre-trained classification system based on a convolutional neural network (CNN) to analyze behavioral features in the image sequence.

[0032] Determine whether the target behavior meets the criteria for suspicious behavior, and further determine the specific behavior category;

[0033] If the system identifies suspicious behavior, it will output the identification results in real time and send an alert to the security or monitoring system.

[0034] Furthermore, when dynamically tracking pedestrian targets and continuously capturing and saving pedestrian images within the anchor frame, the YOLO (You Only Look Once: Unified, Real-Time Object Detection) target detection algorithm is used.

[0035] Furthermore, when performing edge detection on the saved image, the Sobel operator or the Canny edge detection algorithm is used.

[0036] Furthermore, the suspicious behavior included climbing over fences.

[0037] A low-cost gait recognition and suspicious behavior detection system, comprising:

[0038] The camera module is used to capture pedestrian targets in the monitored area in real time.

[0039] The target detection module is used to dynamically track targets and continuously capture and save pedestrian images within the anchor frame;

[0040] The image processing module is used to perform edge detection processing on the saved images and to select N adjacent consecutive images for stacking;

[0041] The behavior recognition module is used to input the stacked images into a pre-trained classification system based on a convolutional neural network (CNN), analyze the behavioral features in the image sequence, determine whether the target behavior meets the suspicious behavior criteria, and further determine the specific behavior category.

[0042] The alarm module is used to output the identification results in real time and send an alarm to the security or monitoring system if the system identifies suspicious behavior.

[0043] The camera module captures images of pedestrian targets in the monitored area in real time;

[0044] The target detection module uses the YOLO algorithm to detect and track pedestrian targets, and captures pedestrian images within the anchor box;

[0045] The image processing module performs edge detection on the captured pedestrian images and extracts image contour information;

[0046] The image stacking module stacks N consecutive edge-detected images to form a three-dimensional tensor;

[0047] The behavior recognition module inputs the stacked images into the CNN model, analyzes the behavioral features, and performs behavior recognition.

[0048] If suspicious behavior is detected, the alarm module will output the identification results in real time and send an alarm to the security or monitoring system.

[0049] Furthermore, the target detection module employs the YOLO target detection algorithm.

[0050] Furthermore, the image processing module employs the Sobel operator or the Canny edge detection algorithm.

[0051] Furthermore, the behavior recognition module employs a pre-trained convolutional neural network model.

[0052] The beneficial effects of this invention are as follows: by combining the YOLO target detection algorithm and a deep learning image classification model, it achieves recognition accuracy and real-time performance similar to LiDAR solutions in a cost-effective and efficient manner. Specific steps include using a camera to capture pedestrian targets in real time, dynamic tracking, and image edge processing, and using image stacking and CNN classification to achieve behavior analysis and automatic alarm output.

[0053] Beneficial effects of the present invention

[0054] (1) Compared to LiDAR-based systems, this solution relies entirely on the existing camera surveillance network, eliminating the need for expensive LiDAR equipment and reducing hardware investment. Furthermore, the system requires no complex installation or maintenance, further reducing labor and maintenance costs. Compared to traditional LiDAR, this solution achieves similar performance at a lower cost, reducing the economic burden of system deployment and operation, making it particularly suitable for large-scale surveillance applications.

[0055] (2) By combining the YOLO algorithm and the CNN classification model, this solution achieves high accuracy and reliability in pedestrian target capture and behavior recognition. The YOLO target detection algorithm can efficiently identify and track pedestrians, while the CNN model analyzes pedestrian behavior through stacked images to accurately judge suspicious behaviors. Compared with traditional gait recognition, this solution can maintain stable recognition performance in complex environments, effectively improving the recognition efficiency and response speed of the monitoring system.

[0056] (3) The solution employs efficient YOLO detection and CNN classification models to process pedestrian behavior, enabling real-time tracking and identification of targets in the monitoring screen, and achieving real-time monitoring and alarm output of suspicious behavior. Compared with traditional methods, this solution significantly improves the system's real-time performance and response speed, providing strong support for the immediate handling of security incidents.

[0057] (4) This solution extracts gait features through edge detection, image stacking, and other techniques, and combines them with a CNN deep learning model to achieve good adaptability to different lighting conditions, occlusion, and other environments, ensuring the effectiveness of behavior recognition in complex scenarios. Compared with LiDAR systems, this solution is more applicable in more scenarios, reduces the false detection rate under specific weather conditions, and thus improves the overall reliability of the system.

[0058] (5) Unlike LiDAR, which captures precise spatial information, this solution does not directly collect the three-dimensional contours of the human body. Instead, it utilizes image data to complete gait recognition and behavior analysis, thus reducing the intrusion on personal privacy. Simultaneously, the data access control and encryption mechanisms incorporated into the system ensure data security and meet the privacy protection requirements of the surveillance field.

[0059] (6) This solution is highly scalable and can combine and optimize various deep learning algorithms according to the needs of different application scenarios. Through modular design, the system can be easily extended to other fields such as identity recognition and security monitoring, providing flexible solutions for different needs.

[0060] Traditional LiDAR-based gait recognition suffers from high equipment costs, poor environmental adaptability, and high privacy risks. This invention significantly reduces costs and privacy risks through a camera-based solution, while maintaining high recognition accuracy and real-time performance in complex environments, effectively addressing the shortcomings of the prior art.

[0061] This invention aims to provide a low-cost, high-precision gait recognition-based method for detecting suspicious behavior. It utilizes a camera system and deep learning to replace the gait recognition function of LiDAR while maintaining equivalent or similar detection performance. Ultimately, this invention achieves the goals of low cost, high efficiency, and privacy protection while also providing greater adaptability, reliability, and real-time performance, meeting the requirements set forth in the invention's objectives.

[0062] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0063] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0064] Figure 1 Example image showing the stacking of records of pedestrians climbing over fences;

[0065] Figure 2 Example image showing stacked records of pedestrian phone calls;

[0066] Figure 3 Example image showing the stacked records of pedestrians' sitting and resting behaviors;

[0067] Figure 4 Example image showing the stacking of pedestrian standing behavior records;

[0068] Figure 5 Example image showing the stacking of pedestrian walking behavior records. Detailed Implementation

[0069] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0070] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0071] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0072] This invention proposes a low-cost gait recognition and suspicious behavior detection system based on cameras. The system utilizes existing camera surveillance networks and combines them with deep learning algorithms to achieve automatic recognition and alerting of pedestrian behavior.

[0073] System Flow:

[0074] 1. Target Detection and Capture: This method utilizes existing surveillance cameras and deep learning-based target recognition algorithms (such as YOLO) to detect and capture pedestrians within the monitored area. YOLO's efficient anchor box mechanism enables real-time identification and labeling of pedestrian targets. The specific principle is as follows:

[0075] 1.1 Image segmentation

[0076] The input image is divided into S×S grids. Each grid is responsible for detecting targets within its coverage area.

[0077] 1.2 Bounding Box (Anchor Box) Prediction

[0078] Each grid cell predicts B bounding boxes. Each box contains the following four coordinates and a confidence score:

[0079] Position coordinates: Center coordinates (x, y) of the box, width w, height h.

[0080] Confidence score C: The confidence score represents the probability that the bounding box contains the target. The formula is as follows:

[0081] C = P(object) × IOU pred,truth

[0082] Where P(object) is the probability that the bounding box contains the object, and IOU pred,truth Intersection over Union (IoU) between predicted bounding boxes and ground truth bounding boxes

[0083] 1.3 Class Probability

[0084] Each grid cell predicts the conditional probability of a target's class, which is the probability that a target in that grid cell belongs to a certain class. For each class m, the probability is:

[0085] P(class m |object)

[0086] 1.4 Output Tensors

[0087] The final output is an S×S×(B×5+C) tensor, where: S×S is the number of grids, B×5 is the parameters of each bounding box (including position, width, height and confidence); and C is the probability of each object class.

[0088] 1.5 Confidence Filtering

[0089] Only retain boxes that meet the C threshold. This step ensures that boxes with low confidence are filtered out, thus reducing interference.

[0090] 1.6 Non-maximum suppression (NMS)

[0091] Duplicate boxes are removed using Non-Maximum Suppression (NMS). For all predicted boxes in each category, the Interchange of Units (IOU) is calculated; if the IOU exceeds a specified threshold, only the box with the highest confidence is retained.

[0092] If IOU pred1,truth2 If the threshold is reached, only boxes with higher confidence levels will be retained.

[0093] Through the above steps, YOLO achieves fast and accurate detection of pedestrian targets in an image during a single forward propagation.

[0094] 2. Dynamic tracking and image capture: After detecting a pedestrian target, the target is dynamically tracked, and the pedestrian image in the anchor frame is continuously captured and saved to ensure the complete movement process of the target is captured.

[0095] 2.1 Calculate the coordinates of the top-left and bottom-right corners of the bounding box.

[0096] Based on the anchor frame information obtained in process 1 (center coordinates (x, y), width w, height h), the top left corner (x... min ,y min ) and the coordinates of the lower right corner (x max ,y max This can be represented as:

[0097]

[0098] 2.2 Extracting image regions:

[0099] In the original image, use the pixel range [x] min :x max ,ymin :y max Image cropping. Assuming the original image matrix is ​​Image, the sub-image within the cropping box can be represented as:

[0100] Image sub =Image[x min :x max ,y min :y max ]

[0101] Using the formula above, we can extract the image region within the bounding box from the original image.

[0102] 3. Edge Detection and Image Stacking: Edge detection is performed on the saved images to extract the main contour information. N adjacent consecutive images are stacked to capture the dynamic features of the target pedestrian on the time axis.

[0103] 3.1 Edge Detection Process

[0104] Edge detection typically uses image gradients or edge detection operators, such as the Sobel operator or the Canny edge detection algorithm. The following description uses the Sobel operator as an example to illustrate the edge detection process for a cropped image.

[0105] Let the cropped image region be Image sub The pixel position is (i,j). Calculate the horizontal direction G at each pixel. x and vertical direction G y The gradient.

[0106]

[0107] The asterisk (*) here indicates a convolution operation.

[0108] Then use G x and G y Calculate the gradient magnitude (edge ​​intensity) G at each pixel:

[0109]

[0110] To refine the edges, non-maximum suppression can be applied to G, retaining only the gradient locations of local maxima as edges.

[0111] By setting a threshold T, pixels with an amplitude G greater than T are retained as edge points, while other pixels are set to 0.

[0112]

[0113] The resulting Edge is a binarized image after edge detection, showing the edges in the cropped image.

[0114] 3.2 Image Stacking

[0115] Stacking N consecutive edge-detected images can be represented as a three-dimensional tensor stacking process. Let the edge detection result of each image be denoted as Edge. k Where k = 1, 2, ..., represents the k-th frame image, the following is a description of the stacking process:

[0116] The edge detection result for each cropped image k is denoted as the matrix Edge. k (i,j) represents the edge information of the image at pixel (i,j).

[0117] N consecutive edge detection images are stacked into a three-dimensional tensor Stack(i,j,k), where the third dimension k represents the order of the time series. Specifically, it is represented as:

[0118] Stack(i,j,k) = Edge k (i,j)

[0119] Final stacking result:

[0120] Stack = {Edge1, Edge2, ..., Edge} N}

[0121] 4. Gait Behavior Classification: The stacked images are input into a pre-trained classification system based on a Convolutional Neural Network (CNN) to analyze the behavioral features in the image sequence. The system uses a classification model to determine whether the target behavior meets the criteria for suspicious behavior and further identifies the specific behavior category. The specific steps are as follows:

[0122] The input tensor stack is fed into the pre-trained CNN model. Assume the CNN's mapping function is f. CNN The forward propagation process is as follows:

[0123] Feature = f CNN (Stack)

[0124] Here, Feature is the feature vector extracted by CNN.

[0125] The feature vector is input into a fully connected layer for classification, assuming the mapping function of the fully connected layer is f. FC And use the softmax activation function to output the class probabilities:

[0126] Class Probabilities=softmax(f FC(Feature)

[0127] Final predicted image category The category corresponding to the maximum value in the category probability:

[0128] Output and Alarm: If the system identifies suspicious behavior, it will output the identification result in real time and send an alarm to the security or monitoring system to facilitate timely response by relevant personnel.

[0129] This invention integrates existing camera monitoring systems and the YOLO algorithm, utilizing a CNN-based image stacking classification method to achieve low-cost, high-efficiency gait recognition and suspicious behavior detection. The entire system features efficient data processing and accurate real-time performance, ensuring stable recognition accuracy in various environments.

[0130] YOLO-based target detection: Using the YOLO algorithm to detect pedestrian targets in real time in the monitoring screen, and combining it with the anchor box mechanism to accurately capture and track pedestrians.

[0131] Image edge processing and stacking: Edge detection and stacking are performed on continuous images to centrally represent pedestrian dynamic features on the time axis, providing key features for subsequent gait recognition and behavior judgment.

[0132] CNN classification model for behavior recognition: Stacked images are input into a pre-trained CNN model for classification, accurately determining whether the target behavior belongs to the "suspicious" category and identifying specific behavior patterns.

[0133] Real-time alert mechanism: Once suspicious behavior is detected, the system will automatically issue an alert and output the results to help monitoring personnel react quickly.

[0134] Efficient use of existing equipment: This invention does not require expensive equipment such as LiDAR, but only uses existing cameras and deep learning algorithms to achieve gait recognition and anomaly detection, effectively reducing costs and simplifying deployment.

[0135] Highly adaptable to multiple scenarios: The system is highly adaptable and can operate stably in a variety of complex environments, making it suitable for fields such as security and public monitoring.

[0136] By implementing the above technical points, this invention can efficiently and economically complete gait recognition and suspicious behavior detection, and has broad application prospects in public safety and intelligent security.

[0137] To evaluate the behavior recognition method proposed in this invention, we used five types of pedestrian behavior records extracted from publicly available videos on the Internet through the methods described in the technical solution and the key points of the invention. These included four normal behaviors (standing, walking, making a phone call, and sitting to rest) and one abnormal behavior (climbing over a fence), and we employed a cross-validation method. Figures 1-5 This demonstrates an example of each behavior (these examples are the result of stacking images after edge detection):

[0138] We selected three CNN models pre-trained using different strategies for behavior recognition testing. The specific differences in training strategies are shown in Table 1.

[0139] Table 1

[0140]

[0141] Finally, the aggregation confusion matrices for behavior recognition in the test scenarios are shown in Tables 2 to 4. Different tables represent the recognition test results of different pre-trained CNN models. The value in the m-th row and n-th column of the matrix represents the number of samples that were identified as the n-th activity in the m-th activity. This table shows the actual accuracy information of this method. Table 2 shows the overall average accuracy of CNN_1 (98.74%); Table 3 shows the overall average accuracy of CNN_2 (97.48%); Table 4 shows the overall average accuracy of CNN_3 (98.11%).

[0142] Table 2

[0143]

[0144] Table 3

[0145]

[0146] Table 4

[0147]

[0148]

[0149] Suspicious behavior categories (fence climbing) and most normal behaviors were almost always correctly identified, while walking behavior received a significant number of false alarms. This is presumably because small hand movements are sometimes difficult to distinguish due to obstruction, leading to confusion between walking and phone calls in some situations. It is also noteworthy that the overall number of false alarms was quite low (less than 5% of genuine behavior), indicating that the system rarely issues unexpected alerts when detecting pedestrian behavior types.

[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A low-cost gait recognition and suspicious behavior detection method, characterized by: Includes the following steps: Use cameras to capture pedestrian targets in real time within the monitored area; The YOLO target detection algorithm is used to dynamically track pedestrian targets, and the pedestrian images in the anchor frame are continuously captured and saved; Edge detection is performed on the saved image to extract its contour information; Select adjacent consecutive N The images are stacked to form a three-dimensional tensor; Continuous N Images that have undergone edge detection are stacked, which can be represented as the superposition process of three-dimensional tensors; the edge detection result of each image is denoted as Edge. k ,in k =1,2,..., N Indicates the first k The frame image stacking process is as follows: For each cropped image k The edge detection results are denoted as a matrix. , indicating the image in pixels Edge information at the location; Will N A series of consecutive edge detection images are stacked into a three-dimensional tensor The third dimension k The order of a time series is represented as follows: The final stacking result is: The stacked images are input into a pre-trained classification system based on a convolutional neural network (CNN) to analyze behavioral features in the image sequence. Determine whether the target behavior meets the criteria for suspicious behavior, and further determine the specific behavior category; If the system identifies suspicious behavior, it will output the identification results in real time and send an alert to the security or monitoring system.

2. The low-cost gait recognition and suspicious behavior detection method according to claim 1, characterized in that: When performing edge detection on the saved image, the Sobel operator or the Canny edge detection algorithm is used.

3. The low-cost gait recognition and suspicious behavior detection method according to claim 1, characterized in that: The suspicious behavior included climbing over fences.

4. A low-cost gait recognition and suspicious behavior detection system, characterized in that: The system includes: The camera module is used to capture pedestrian targets in the monitored area in real time. The target detection module is used to dynamically track targets and continuously capture and save pedestrian images within the anchor frame; The image processing module is used to perform edge detection processing on the saved images and to select N adjacent consecutive images for stacking; The behavior recognition module is used to input the stacked images into a pre-trained classification system based on a convolutional neural network (CNN), analyze the behavioral features in the image sequence, determine whether the target behavior meets the suspicious behavior criteria, and further determine the specific behavior category. The alarm module is used to output the identification results in real time and send an alarm to the security or monitoring system if the system identifies suspicious behavior. The camera module captures images of pedestrian targets in the monitored area in real time; The target detection module uses the YOLO algorithm to detect and track pedestrian targets, and captures pedestrian images within the anchor box; The image processing module performs edge detection on the captured pedestrian images and extracts image contour information; The image stacking module stacks N consecutive edge-detected images to form a three-dimensional tensor; it then stacks the consecutive... N Images that have undergone edge detection are stacked, which can be represented as the superposition process of three-dimensional tensors; the edge detection result of each image is denoted as Edge. k ,in k =1,2,..., N Indicates the first k The frame image stacking process is as follows: For each cropped image k The edge detection results are denoted as a matrix. , indicating the image in pixels Edge information at the location; Will N A series of consecutive edge detection images are stacked into a three-dimensional tensor The third dimension k The order of a time series is represented as follows: The final stacking result is: The behavior recognition module inputs the stacked images into the CNN model, analyzes the behavioral features, and performs behavior recognition. If suspicious behavior is detected, the alarm module will output the identification results in real time and send an alarm to the security or monitoring system.

5. The low-cost gait recognition and suspicious behavior detection system according to claim 4, characterized in that: The image processing module employs either the Sobel operator or the Canny edge detection algorithm.

6. The low-cost gait recognition and suspicious behavior detection system according to claim 4, characterized in that: The behavior recognition module uses a pre-trained convolutional neural network model.

Images