Table tennis tactical analysis method based on lightweight edge features and grading labeling

Abstract

The invention relates to a table tennis tactical analysis method based on lightweight edge features and hierarchical annotation, and the method comprises the steps: deploying a lightweight CNN model at an edge device, carrying out the calculation of video data, making full use of the calculation capability of the edge device, and reducing the calculation burden of a cloud end; through an edge device-cloud task grading cooperation mechanism, the table tennis tactical analysis efficiency is greatly improved; meanwhile, based on a task grading mechanism, the recognition accuracy of the second-level task and the third-level task is refined and upgraded through the cloud, and the precision of table tennis tactical analysis is guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sports big data analysis and edge computing, in particular to a table tennis tactics analysis method based on light edge features and hierarchical labeling. BACKGROUND

[0002] Table tennis is one of the most popular sports in the world. It is loved by people of all ages in daily life because of its low entry threshold and strong competitive interest. In professional table tennis matches, players use their quick reactions, rapid attacks, precise controls, and solid defenses to bring the audience one thrilling match after another. Although table tennis and other sports appear to be similar, they mainly test the physical fitness of athletes, such as eye-brain-hand synchronization, lower body stability and flexibility, wrist power explosion, etc. However, unlike other sports, table tennis tests athletes' use of techniques and tactics, i.e., how to use techniques to counterattack and where to hit the ball to gain an advantage in different situations.

[0003] Previously, many table tennis team analysis methods used manual annotation by analysts marking the ball hitting type (such as loop, smash, and block), landing area, and paddle angle frame by frame. The drawbacks of this method are: Low efficiency: It takes several hours to several tens of hours to annotate a match video, which cannot meet the real-time analysis requirements; Strong subjectivity: Different analysts have different standards for judging the same action, resulting in poor consistency of the annotation results; Data lag: The long processing period makes it difficult to achieve immediate feedback during training. In recent years, some studies have attempted to use deep learning models (such as ResNet and YOLO) to automatically extract and label table tennis video features in the cloud. The specific process is as follows: the camera collects the video and uploads it to the cloud server, runs the complex neural network model on the high-performance computing device, and outputs the labeling results. The disadvantages of this method include: High computing cost: The cloud server needs to handle a large amount of video data transmission and complex model calculation, resulting in significant increases in hardware investment and energy consumption costs; Poor real-time performance: There is a network delay (usually 200ms-500ms) between uploading the video to the cloud and returning the results, which cannot meet the real-time tactics analysis requirements on site; Waste of edge device resources: The local computing power of edge computing terminals (such as sports cameras and smart set-top boxes) is not fully utilized, and the centralized processing architecture is relied on. The existing method depends on manpower, resulting in low efficiency, or depends on the cloud, resulting in high cost and low real-time performance, and cannot realize lightweight and efficient table tennis skill and tactic feature extraction and labeling on the edge device. SUMMARY

[0005] The technical problem to be solved by the present application is how to provide a lightweight feature extraction suitable for edge devices to realize high-precision and efficient table tennis skill and tactic labeling analysis.

[0006] The present application provides a table tennis tactic analysis method based on lightweight edge features and hierarchical labeling, based on edge devices, the cloud and terminals, comprising: Step 1, the edge device collects a table tennis video stream, and pre-processes to obtain a table area image sequence; Step 2, deploying a lightweight CNN model on the edge device, performing multi-scale target detection on the input table area image sequence, and outputting a feature vector containing the type of hitting, the trajectory of the ball and the coordinates of the landing point; Step 3, calculating the hitting type confidence based on the hitting type, and classifying the first-level task, the second-level task and the third-level task according to the hitting type confidence as the hierarchical standard; If it is a first-level task, the edge device labels the corresponding position of the table tennis video stream according to the feature vector and outputs it to the terminal display; If it is a second-level task, the edge device uploads the feature vector to the cloud, and the cloud performs refinement and returns it to the edge device, and the edge device labels the corresponding position of the table tennis video stream based on the cloud refinement result and the feature vector and outputs it to the terminal display; If it is a third-level task, the edge device uploads the feature vector to the cloud, and the cloud performs tactic combination recognition and returns it to the edge device, and the edge device labels the corresponding position of the table tennis video stream based on the tactic combination recognition result and outputs it to the terminal display.

[0007] Compared with the prior art, the present application has the following advantages: the present application deploys a lightweight CNN model on the edge device to calculate video data, fully utilizes the computing power of the edge device, and reduces the computing burden of the cloud; through the edge device-cloud task hierarchical collaborative mechanism, the efficiency of table tennis tactic analysis is greatly improved. At the same time, based on the task hierarchical mechanism, the recognition accuracy of the second-level task and the third-level task is refined and upgraded by the cloud, ensuring the accuracy of table tennis tactic analysis.

[0008] In one possible implementation, the pre-processing in step 1 includes: Step 101, using an adaptive frame rate detection algorithm to increase the sampling frame rate in the fast hitting stage of the table tennis video stream and reduce the sampling frame rate in the non-hitting stage; sampling to obtain a continuous frame image sequence; Step 102, the background difference method is used to remove the interference of the static scene in the image, and the region of interest is focused on the billiard table region to obtain a billiard table region image sequence.

[0009] In a possible implementation, the lightweight CNN model constructed in step 2 includes: An input layer is constructed to accept the billiard table region image sequence obtained by preprocessing; A feature extraction layer is constructed, including two deep separable convolution units connected in sequence, each deep separable convolution unit adopts the lightweight bottleneck structure of MobileNetv3, and an SE attention mechanism is introduced into the lightweight bottleneck structure of MobileNetv3 to enhance the target feature region, and the shallow feature map output by the first deep separable convolution unit is input into the second deep separable convolution unit to obtain a deep feature map; A feature fusion layer is constructed to skip connect the first deep separable convolution unit and the second deep separable convolution unit, and the shallow feature map and the deep feature map are spliced and fused by a feature network pyramid to obtain a feature fusion map; A multi-branch output head is constructed, including a hitting type output head, a ball route trajectory output head, and a landing point coordinate output head; The hitting type output head performs global average pooling on the feature fusion map to output a hitting type feature vector; The ball route trajectory output head processes the feature fusion map through three fully connected layers to output a coordinate sequence to form a ball route trajectory; The landing point coordinate output head processes the feature fusion map through two fully connected layers to obtain a landing point coordinate; The constructed input layer, feature extraction layer, feature fusion layer, and multi-task output head are subjected to quantization compression, structure pruning, and knowledge distillation optimization processing to obtain a lightweight CNN model.

[0010] In a possible implementation, the calculation formula for calculating the hitting type confidence in step 3 is: ; In the formula, represents the hitting type threshold, represents the total number of categories; represents an exponential transformation for calculating the confidence.

[0011] In a possible implementation, the classification of the first-level task, the second-level task, and the third-level task based on the hitting type confidence in step 3 specifically includes: If the hitting type confidence value is greater than or equal to 90%, it is a first-level task; If the execution degree value of the ball hitting type is in the range of [60%~90%], it is a secondary task; Or the execution degree value of the ball hitting type is less than 60%, it is a tertiary task.

[0012] In a possible implementation, the processing procedure of the primary task in step 3 includes: Step A1, the edge device labels the feature vector output by the lightweight CNN model at the corresponding position of the table tennis video stream, and outputs it to the terminal display.

[0013] In a possible implementation, the processing procedure of the secondary task in step 3 includes: Step B1, the edge device uploads the characteristic vector output by the lightweight CNN model to the cloud, and the cloud calculates the ball speed of the table tennis and returns it to the edge device; Step B2, the edge device labels the characteristic vector output by the lightweight CNN model and the ball speed at the corresponding position of the table tennis video stream, and outputs it to the terminal display.

[0014] In a possible implementation, the processing procedure of the tertiary task in step 3 includes: Step C1, the edge device uploads the characteristic vector output by the lightweight CNN model to the cloud; Step C2, the cloud calculates the ball speed of the table tennis; Step C3, the cloud calls a tactical combination feature library containing a number of tactical combinations of ball hitting types, landing point coordinates, and ball speed intervals, performs tactical combination recognition based on the tactical feature library, and returns the recognition result to the edge device; Step C4, the edge device labels the recognized tactical combination at the corresponding position of the table tennis video stream, and outputs it to the terminal display.

[0015] In a possible implementation, the cloud calculating the ball speed of the table tennis includes: The cloud performs deep mining on the characteristic vector through the deployed YoLoV5 neural network model to obtain the pixel displacement of the table tennis between adjacent frames: ; Obtain the camera frame rate , calculate the interval between adjacent frames ; Calculate the normalized pixel resolution ; in the formula, represents the pixel length of the table in the table area image; Calculate the ball speed of the table tennis in the x direction and the y direction: ; ; Computing the ball speed of a table tennis ball . BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The flowchart of the present application is shown. DETAILED DESCRIPTION

[0017] First of all, those skilled in the art should understand that these embodiments are only used to explain the technical principles of the embodiments of the present application, and are not intended to limit the protection scope of the embodiments of the present application. Those skilled in the art can adjust them as needed in order to adapt to specific application occasions.

[0018] In the description of the embodiments of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "connected", "connected" should be understood in a broad sense, for example, it can be fixedly connected, or it can be detachably connected, or it can be integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium. For those skilled in the art, the specific meaning of the above terms in the embodiments of the present application can be understood according to the specific circumstances.

[0019] In the embodiments of the present application, unless otherwise explicitly specified and limited, the first feature is "on" or "under" the second feature. The first and second features can be in direct contact, or the first and second features can be in indirect contact through an intermediate medium. Moreover, the first feature "above", "above" and "above" the second feature can be that the first feature is directly above or obliquely above the second feature, or it only means that the horizontal height of the first feature is higher than that of the second feature. The first feature "below", "below" and "below" the second feature can be that the first feature is directly below or obliquely below the second feature, or it only means that the horizontal height of the first feature is less than that of the second feature.

[0020] The present application will be further described in detail below in combination with the drawings and specific embodiments.

[0021] Referring to Figure 1 The embodiments of the present application disclose a table tennis tactics analysis method based on lightweight edge features and hierarchical labeling, based on edge devices, clouds and terminals, including: Step 1, the edge device collects the table tennis video stream, and the edge device in the embodiments of the present application is, for example, a camera machine built-in chip or an edge computing box; then, the table tennis video stream is preprocessed to obtain a table image sequence; specifically including: Step 101, in order to reduce the data processing amount, an adaptive frame rate detection algorithm is used to increase the sampling frame rate in the fast ball hitting stage of the table tennis video stream and to reduce the sampling frame rate in the non-ball hitting stage; a continuous frame image sequence is obtained by sampling; in the embodiment of the application, the sampling frame rate is increased to 60fps in the fast ball hitting stage (such as smash and fast attack), and is reduced to 30fps in the non-ball hitting stage; thereby reducing the overall calculation amount of the edge device and improving the processing speed of the edge device; the expression is: ; In the formula, represents the sampling frame rate after dynamic frame skipping, represents a preset motion intensity threshold, in the embodiment of the application, 1000 frames of labeled data (including 500 frames of ball hitting actions and 500 frames of non-ball hitting actions) are used to train a classification model, and the optimal ; represents that the motion intensity value in the table tennis video stream is calculated by the background difference method, and in the embodiment of the application, the ball hitting intensity value is used to adapt to the fast ball hitting stage or the non-ball hitting stage in the table tennis video stream; ; Step 102, the background difference method is used to remove the interference of the static scene in the image, improve the target detection accuracy, focus the region of interest on the table area, obtain the table area image sequence, and also reduce the CPU load of the edge device.

[0022] Step 2, a lightweight CNN model is deployed on the edge device to perform multi-scale target detection on the input table area image sequence, and output a feature vector containing the ball hitting type, the ball route trajectory and the landing point coordinates; wherein, The constructed lightweight CNN model includes: An input layer is constructed to accept the table area image sequence obtained by preprocessing; A feature extraction layer is constructed, including two layers of deep separable convolution units connected in sequence, each deep separable convolution unit adopts the lightweight bottleneck structure of MobileNetv3, and an SE attention mechanism is introduced into the lightweight bottleneck structure of MobileNetv3 to enhance the target feature area, and the shallow feature map output by the first layer of deep separable convolution unit is input into the second deep separable convolution unit to obtain a deep feature map; A feature fusion layer is constructed, which is connected to the first layer of deep separable convolution unit and the second layer of deep separable convolution unit, and the shallow feature map and the deep feature map are spliced and fused by a feature network pyramid to obtain a feature fusion map; A multi-branch output head is constructed, including a ball hitting type output head, a ball route trajectory output head and a landing point coordinate output head; The ball hitting type output head performs global average pooling on the feature fusion graph to output a ball hitting type feature vector; The ball route trajectory output head processes the feature fusion graph through three fully connected layers to output a coordinate sequence to form a ball route trajectory; The landing point coordinate output head processes the feature fusion graph through two fully connected layers to obtain a landing point coordinate; The constructed input layer, feature extraction layer, feature fusion layer, and multi-task output head are subjected to quantization compression, structure pruning, and knowledge distillation optimization processing to obtain a lightweight CNN model.

[0023] The quantization compression of the embodiment of the application specifically includes: 8-bit fixed-point quantization: after training, the weights (W) and activation values (A) are statically quantized (Scale / ZeroPoint calibration), and 32-bit floating-point numbers (FP32) are converted into 8-bit integers (INT8), reducing the amount of calculation (matrix multiplication is accelerated from FP32→INT8); Quantization-aware training (QAT): simulate quantization errors (insert pseudo-quantization nodes) during training to avoid a sharp drop in accuracy after quantization.

[0024] The structure pruning in the embodiment of the application specifically includes: Unstructured pruning: based on weight amplitude sorting, remove redundant filters (such as Pointwise Conv weights of deep separable convolution channels close to 0) smaller than a threshold value; Fine-tuning recovery: after pruning, the model is subjected to lightweight fine-tuning (10 rounds) to recover the accuracy lost due to pruning.

[0025] The knowledge distillation in the embodiment of the application specifically includes: Teacher model: select ResNet50 (pre-trained on the table tennis dataset), output containing soft labels (classification probability distribution) and intermediate layer features; Student model (PingNet-Lite): the loss function includes: Task loss (classification cross-entropy + trajectory / landing point smooth L1); The distillation loss includes soft label cross-entropy and hard label cross-entropy, and the calculation formula is: ; In the formula, represents the distillation weight, ; Then, the lightweight CNN model is optimized by TensorRT to generate an inference engine with FP16 precision, which is deployed to an edge device, and the actual measured single-frame processing time is ≤15ms (@1080p resolution).

[0026] Through the guidance of the teacher model, the lightweight CNN model quickly learns complex features under small samples, and the accuracy is improved.

[0027] The lightweight CNN model of the embodiment of the application realizes lightweight deployment while ensuring the analysis accuracy of the table tennis image sequence through depth separable convolution, multi-scale feature fusion and multi-dimensional compression strategies, and is suitable for real-time processing of edge devices.

[0028] Further, the lightweight CNN model of the embodiment of the application specifically identifies the type of the stroke through the landing point coordinates of the table tennis on the table, the features of the front and back of the racket, and the calculated ball speed; for example: Stroke type: forehand smash, table tennis landing probability > 70% on the right half of the table, of which the middle zone (x=0.4~0.6m, y=0.3~0.5m) accounts for 55%, and the long ball zone (y>0.6m) accounts for 20%.

[0029] Stroke type: backhand push block, table tennis landing probability > 80% on the left half of the table near the net (x=0.1~0.3m, y=0.1~0.2m), and the ball speed is mostly 10~15km / h.

[0030] Stroke type: forehand loop, full-table long ball (y>0.6m) probability > 90%, diagonal landing point (x>0.5m or x<0.2m) accounts for 65%.

[0031] Stroke type: |wiping ball| near-net short ball (y<0.2m), landing point is scattered.

[0032] Stroke type: near-net short ball (y<0.2m), landing point is concentrated.

[0033] The calculation formula of the ball speed in the lightweight CNN model is: Extract the table tennis contour coordinates from the continuous two frames of table region images And , the calculation formula of the ball speed is: ; In the formula, denotes the time interval.

[0034] Step 3, calculate the stroke type confidence based on the stroke type, and classify the first-level task, the second-level task and the third-level task according to the stroke type confidence as the grading standard; wherein, The calculation formula of the stroke type confidence is: ; In the formula, denotes the threshold value of the stroke type , and denotes the total number of categories. representing an exponential transformation of the calculation confidence; If the ball hitting type confidence value is greater than or equal to 90%, it is a first-level task; the edge device labels the corresponding position of the table tennis video stream according to the feature vector and outputs to the terminal display; the specific processing process includes: Step A1, the edge device labels the feature vector output by the lightweight CNN model in the corresponding position of the table tennis video stream and outputs to the terminal display.

[0035] If the ball hitting type execution value is in the range of [60%~90%], it is a second-level task, the edge device uploads the feature vector to the cloud, and the cloud performs refinement and returns to the edge device, and the edge device labels the corresponding position of the table tennis video stream based on the cloud refinement result and the feature vector and outputs to the terminal display; the specific processing process includes: Step B1, the edge device uploads the feature vector output by the lightweight CNN model to the cloud, and the cloud calculates the ball speed and returns to the edge device; Step B2, the edge device labels the feature vector output by the lightweight CNN model and the ball speed in the corresponding position of the table tennis video stream and outputs to the terminal display.

[0036] If the ball hitting type execution value is less than 60%, it is a third-level task, the edge uploads the feature vector to the cloud, and the cloud performs tactical combination recognition and returns to the edge device, and the edge device labels the corresponding position of the table tennis video stream based on the tactical combination recognition result and outputs to the terminal display, the specific processing process includes: Step C1, the edge device uploads the feature vector output by the lightweight CNN model to the cloud; Step C2, the cloud calculates the ball speed of table tennis; Step C3, the cloud calls the tactical combination feature library containing the ball hitting type, landing point coordinates, and ball speed interval several tactical combinations; and performs tactical combination recognition based on the tactical feature library, and returns the recognition result to the edge device; Step C4, the edge device labels the corresponding position of the table tennis video stream based on the recognized tactical combination and outputs to the terminal display.

[0037] In addition, the embodiment of the application establishes a task queue in the cloud, and preferentially processes the second-level task (response time ≤100ms), and the third-level task is processed in batches during the non-peak period; Further, the cloud calculating the ball speed of table tennis includes: The cloud performs deep mining on the feature vector through the deployed YoLoV5 neural network model to obtain the pixel displacement of table tennis between adjacent frames: ; Acquiring a camera frame rate , calculating a neighboring frame interval ; calculating a normalized pixel resolution ; wherein, represents a pixel length of the table in the table region image; calculating the ball speed of the table tennis in x direction and y direction ; ; calculating the ball speed of the table tennis .

[0038] In the description of the embodiments of the present application, it should be noted that in the description of the present application, the terms indicating the direction or position relationship such as "inner", "outer" and the like are based on the direction or position relationship shown in the drawings, which is merely for the convenience of description, and does not indicate or imply that the device or member must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application.

[0039] In the description of the present application, the description of the terms "one embodiment", "some embodiments", "in this embodiment", "specific example", or "some examples" and the like means that the specific features, mechanisms, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, mechanisms, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. Furthermore, the skilled person in the art can combine and combine the different embodiments or examples described in the specification and the features of the different embodiments or examples without contradiction.

[0040] The above is merely a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A table tennis tactical analysis method based on lightweight edge features and hierarchical annotation, based on edge devices, cloud, and terminal, characterized in that, include: Step 1: The edge device acquires a video stream of a ping-pong ball and preprocesses it to obtain an image sequence of the table area; Step 2: Deploy a lightweight CNN model on an edge device to perform multi-scale target detection on the input image sequence of the table area, and output a feature vector containing the shot type, ball trajectory and landing point coordinates; Step 3: Calculate the confidence level of the shot type based on the shot type, and classify the tasks into Level 1, Level 2 and Level 3 based on the confidence level of the shot type. If it is a Level 1 task, the edge device will annotate the corresponding positions of the ping-pong video stream according to the feature vector and output it to the terminal for display; If it is a level 2 task, the edge device uploads the feature vector to the cloud, which refines it and returns it to the edge device. The edge device annotates the corresponding position of the table tennis video stream based on the refinement result and feature vector from the cloud, and outputs it to the terminal for display. If it is a level 3 task, the edge device will upload the feature vector to the cloud, and the cloud will perform tactical combination recognition and return it to the edge device. The edge device will then mark the corresponding position of the table tennis video stream based on the tactical combination recognition result and output it to the terminal for display.

2. The table tennis tactical analysis method based on lightweight edge features and hierarchical annotation according to claim 1, characterized in that, The preprocessing in step 1 includes: Step 101: An adaptive frame rate detection algorithm is used to increase the sampling frame rate during the fast hitting phase of the table tennis video stream and decrease the sampling frame rate during the non-hitting phase; a continuous frame image sequence is obtained by sampling. Step 102: Use background subtraction to remove interference from static scenes in the image, focus the region of interest on the table area, and obtain the image sequence of the table area.

3. The table tennis tactical analysis method based on lightweight edge features and hierarchical annotation according to claim 1, characterized in that, The lightweight CNN model constructed in step 2 includes: Construct an input layer to receive the preprocessed image sequence of the pool table area; A feature extraction layer is constructed, comprising two layers of depthwise separable convolutional units connected in sequence. Each depthwise separable convolutional unit adopts the lightweight bottleneck structure of MobileNetv3, and an SE attention mechanism is introduced into the lightweight bottleneck structure of MobileNetv3 to enhance the target feature region. The shallow feature map output by the first depthwise separable convolutional unit is input into the output of the second depthwise separable convolutional unit to obtain the deep feature map. A feature fusion layer is constructed, which is a jump connection between the first layer of depthwise separable convolutional units and the second layer of depthwise separable convolutional units. The shallow feature maps and deep feature maps are spliced ​​and fused using a feature network pyramid to obtain the feature fusion map. Construct a multi-branch output head, including a shot type output head, a ball trajectory output head, and a landing point coordinate output head; The shot type output head performs global average pooling on the feature fusion map to output a shot type feature vector; The ball trajectory output head processes the feature fusion map through three fully connected layers and outputs a coordinate sequence to form the ball trajectory. The landing point coordinate output head processes the feature fusion map through two fully connected layers to obtain the landing point coordinates; The constructed input layer, feature extraction layer, feature fusion layer, and multi-task output head are optimized by quantization compression, structural pruning, and knowledge distillation to obtain a lightweight CNN model.

4. The table tennis tactical analysis method based on lightweight edge features and hierarchical annotation according to claim 1, characterized in that, The formula for calculating the confidence level of the shot type based on the shot type in step 3 is as follows: ； In the formula, Indicates the type of shot The threshold, Indicates the total number of categories; This represents the exponential transformation performed to calculate the confidence level.

5. The table tennis tactical analysis method based on lightweight edge features and hierarchical annotation according to claim 1, characterized in that, Step 3, which classifies tasks into Level 1, Level 2, and Level 3 based on confidence level of shot type, specifically includes: If the confidence value of the shot type is greater than or equal to 90%, it is a Level 1 task; If the execution rate of the shot type is within the range of [60%~90%], it is a level 2 task; If the execution rate of the shot type is less than 60%, it is a level three task.

6. The table tennis tactical analysis method based on lightweight edge features and hierarchical annotation according to claim 1, characterized in that, The processing procedure for the first-level task in step 3 includes: Step A1: The edge device labels the feature vectors output by the lightweight CNN model at the corresponding positions in the ping-pong video stream and outputs them to the terminal for display.

7. The table tennis tactical analysis method based on lightweight edge features and hierarchical annotation according to claim 1, characterized in that, The processing procedure for the secondary task in step 3 includes: Step B1: The edge device uploads the feature vector output by the lightweight CNN model to the cloud, and the cloud calculates the ball speed of the ping-pong ball and returns it to the edge device. In step B2, the edge device labels the feature vector and ball speed output by the lightweight CNN model at the corresponding positions in the ping-pong video stream and outputs them to the terminal for display.

8. The table tennis tactical analysis method based on lightweight edge features and hierarchical annotation according to claim 1, characterized in that, The processing procedure for the third-level task in step 3 includes: Step C1: The edge device uploads the feature vector output by the lightweight CNN model to the cloud; Step C2, the cloud calculates the ball speed of the ping-pong ball; Step C3: The cloud retrieves a tactical combination feature library containing several tactical combinations of shot type, landing point coordinates, and ball speed range; and performs tactical combination identification based on the tactical feature library, and returns the identification results to the edge device; In step C4, the edge device annotates the corresponding positions in the table tennis video stream based on the identified tactical combinations and outputs the annotations to the terminal for display.

9. The table tennis tactical analysis method based on lightweight edge features and hierarchical annotation according to claim 7 or 8, characterized in that, The cloud-based calculation of the ball speed in ping-pong includes: The cloud-based system uses a deployed YoLoV5 neural network model to perform deep mining of feature vectors, obtaining the pixel displacement of the ping-pong ball between adjacent frames: ; Get camera frame rate Calculate the interval between adjacent frames ; Calculate the normalized pixel resolution In the formula, This indicates the pixel length of the pool table in the image of the pool table area; Calculate the ball speed in the x and y directions of a ping-pong ball: ； ； Calculate the speed of a table tennis ball .

Images