Action normativity real-time evaluation method and system based on edge calculation and dynamic weight analysis

Abstract

The invention discloses an action normalization real-time evaluation method and system based on edge calculation and dynamic weight analysis. According to the method, video streams are collected through an edge computing terminal, human skeleton key points are extracted, and dual modes of teaching input and real-time evaluation are supported. In a teaching mode, the system automatically generates a dynamic weight matrix focused on a key action part by analyzing a joint point displacement variance of a standard action; and in a real-time evaluation mode, performing millisecond judgment on the action by utilizing a multi-stage cascade filtering mechanism and a geometric feature normalization algorithm in combination with the dynamic weight matrix. The problems that a traditional visual monitoring scheme depends on a high-computing-power cloud server, the privacy leakage risk is large, and the fixed rule adaptability is poor are solved, high-precision, low-delay and non-contact type action normative monitoring on low-power-consumption equipment is achieved, and the method is suitable for the fields of industrial safety production, rehabilitation training, physical examination and the like.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and artificial intelligence, specifically to a technology for monitoring the standardization of human movements in resource-constrained environments. Background Technology

[0002] In scenarios such as industrial safety training, rehabilitation medicine, and sports education, assessing the standardization of human movements is a core requirement. Traditional assessment methods mainly rely on manual supervision or contact sensors. Manual supervision is costly, inefficient, and difficult to standardize; contact sensors require wearable devices, resulting in a poor user experience and high maintenance costs. In recent years, computer vision-based action recognition technology has developed, but existing technologies typically rely on high-performance GPU servers for deep learning inference, requiring video streams to be uploaded to the cloud for processing. This approach not only demands high network bandwidth, leading to high latency and failing to meet real-time error correction needs, but also poses a serious risk of data privacy leaks. Furthermore, most existing algorithms use fixed threshold judgments or general action classification models, lacking adaptability for specific individuals or specific actions. When the application scenario changes (such as changes in shooting angle or differences in user body size), the recognition accuracy drops significantly. Summary of the Invention

[0003] This invention aims to solve the above problems by providing a real-time evaluation method and system for action standardization based on edge computing and dynamic weight analysis.

[0004] This invention achieves high-precision, low-latency, and non-contact motion conformity monitoring on low-power devices by deploying a lightweight attitude estimation model at the edge, combined with a dynamic weight matrix generated by teaching input and a multi-level cascaded filtering mechanism, while effectively protecting user privacy. Attached Figure Description

[0005] Figure 1 This is a system architecture block diagram of an embodiment of the present invention;

[0006] Figure 2 This is a flowchart of the core method in an embodiment of the present invention;

[0007] Figure 3 This is a schematic diagram illustrating the geometric feature calculation principle in an embodiment of the present invention. Detailed Implementation

[0008] The present invention will now be described in detail with reference to the accompanying drawings.

[0009] Example 1: System architecture as follows Figure 1As shown, this embodiment provides a real-time motion compliance evaluation system, mainly comprising an interaction layer and an edge computing terminal. The interaction layer includes an image acquisition device 110 (such as a USB camera) and an interaction feedback device 120 (such as a display screen or speaker). The core protection entity is the edge computing terminal 100, which internally deploys a data preprocessing module 201, a lightweight pose estimation module 202 (accelerated by NPU), a local feature database 203, and a core algorithm unit 300. The core algorithm unit 300 further includes an environment adaptive calibration unit 301, a geometric feature normalization unit 302, and a multi-level constraint determination engine 303. This architecture design ensures that all raw video streams are processed locally and not uploaded to the cloud, achieving privacy protection and low-latency response.

[0010] Example 2: Evaluation method and process as follows Figure 2 As shown, the method flow of the present invention includes the following steps: Step S1: Acquire video stream. The edge terminal reads video frames with a resolution of 1920x1080 or lower in real time. Step S2: Skeleton extraction. The three-dimensional coordinates of 33 key points of the human body are extracted using a lightweight neural network model. Step S3: Determine the current mode. (1) Teaching input mode: In this mode, the system collects the full-cycle action sequence of the standard demonstrator. The processor calculates the displacement variance of each joint on the time axis. If the displacement variance of a certain joint (such as the knee when squatting) is large, it indicates that it is the core feature point of the action, and the system automatically assigns it a higher weight coefficient; otherwise, it assigns a low weight. Finally, a dynamic weight matrix based on variance is generated and stored in the local rule base. This allows the system to adapt to new actions without retraining the model. (2) Real-time evaluation mode: In this mode, the system loads the above dynamic weight matrix and enters a multi-level cascaded filtering process: First-level coarse screening: Check the confidence of key points. If the occlusion is severe, the current frame is directly discarded (saving computing power). Secondary fine screening: Calculates normalized geometric features. Tertiary core judgment: Performs a weighted comparison between real-time features and standard features. If the weighted distance is less than a threshold, it is judged as compliant; otherwise, error correction feedback is generated.

[0011] Example 3: Feature Calculation Principle Figure 3 As shown, to eliminate the influence of user height, body shape, and shooting distance, this invention employs feature normalization processing. The system first performs spatial coordinate calibration on the original skeletal coordinates to correct distortion caused by the camera's perspective. Then, it performs vectorization processing to calculate the angles (Angle θ) of key joints and the relative distance ratios (Ratio r) between joints. These geometric features are dimensionless, enabling standardized evaluation across individuals. Finally, through temporal state mapping, the final action score is output.

Claims

1. A real-time evaluation method for action standardization based on edge computing and dynamic weight analysis, characterized in that, Includes the following steps: Step S1: Acquire real-time video stream data of the target object using the image acquisition unit deployed on the edge computing terminal; Step S2: Process each frame of image using a lightweight human pose estimation model to extract three-dimensional spatial coordinate data containing multiple joint points; Step S3: Determine the current system working mode. If it is the teaching and input mode, proceed to step S4; if it is the real-time evaluation mode, proceed to step S5. Step S4: Collect full-cycle time-series data of the standard action, traverse and calculate the displacement feature matrix of each joint point within the action cycle, generate a dynamic weight matrix based on the statistical distribution characteristics of each element in the displacement feature matrix, and store it in the local rule base; Step S5: Load the dynamic weight matrix, obtain the joint coordinates of the current action, calculate the weighted action standardization score in combination with the preset geometric feature constraints, and output the evaluation result.

2. The method according to claim 1, characterized in that, The specific method for generating the dynamic weight matrix in step S4 includes: calculating the spatial displacement variance of each joint point within the action cycle; determining joint points with variance greater than a preset threshold as action-sensitive points and assigning them higher weight coefficients; determining joint points with variance less than the preset threshold as non-sensitive points and assigning them lower weight coefficients; and normalizing the weight coefficients of all joint points to construct the dynamic weight matrix.

3. The method according to claim 1, characterized in that, Before calculating the action standardization score in step S5, a multi-level cascaded filtering step is also included: First-level filtering: Detect the overall confidence of the human skeleton key points in the current frame. If it is lower than the preset threshold, the current frame is discarded and no further calculation is performed; Second-level filtering: Extract the spatial relative position of key feature points. If it does not meet the preset human biomechanical constraint range, it is determined to be an invalid posture. Third-level judgment: Only if the first two levels of filtering have passed, a high-precision comparison calculation based on the dynamic weight matrix is ​​performed.

4. The method according to claim 1, characterized in that, The comparison calculation in step S5 adopts a geometric feature normalization method: select three associated key points to construct a spatial vector, and calculate the angle feature value between the vectors; calculate the relative distance ratio between key points; compare the angle feature value and the relative distance ratio with the standard threshold range in the local rule base. The comparison process does not depend on the absolute pixel position of the key points in the image coordinate system.

5. The method according to claim 1, characterized in that, After extracting coordinate data in step S2, an environment adaptive calibration step is also included: identifying environmental reference objects or ground planes in the image; constructing a virtual three-dimensional coordinate system and mapping the extracted skeletal key point coordinates to the standard frontal view perspective to eliminate perspective distortion caused by the shooting angle.

6. A real-time evaluation system for motion standardization, characterized in that, include: The image acquisition module is configured to acquire real-time video streams; An edge computing processing module, configured to run the method as described in any one of claims 1 to 5; The storage module is used to store the local rule base and dynamic weight matrix; The interactive feedback module is used to output prompts based on the evaluation results; The system completes the entire calculation process locally on the edge computing terminal without uploading the original video stream to the cloud server.

Images