Automatic recognition method of landing moment in standing long jump based on optical flow analysis

Abstract

The present application relates to a method for automatically identifying the landing moment of standing long jump based on optical flow analysis, which comprises the following steps: pre-processing a high frame rate video sequence in sequence to obtain a high frame rate video optimization sequence; calculating a pixel-level dynamic optical flow field sequence between adjacent frames of the high frame rate video optimization sequence; recursively decomposing the pixel-level dynamic optical flow field sequence to obtain motion components; calculating a complexity index time series corresponding to the motion components; identifying key motion features based on the complexity index time series; processing key motion features based on a multi-level heuristic decision tree model to identify the landing moment frame; constructing an environmental compensation model based on environmental factors and the pixel-level dynamic optical flow field sequence; processing potential landing moment frames and pixel-level dynamic optical flow field sequences based on the environmental compensation model to obtain landing key frames, and calculating the student's long jump score based on the landing key frames, thereby achieving efficient automation throughout the process, comprehensively and meticulously analyzing the movements, and achieving accurate and objective measurement of the scores.

Description

Technical Field

[0001] The present application relates to the technical field of data recognition, and in particular to a method for automatically recognizing the landing moment of a standing long jump based on optical flow analysis. Background Art

[0002] The key to accurately measuring standing long jump results is to accurately capture the moment when students land. Traditional measurement methods mainly rely on manual observation and judgment, which has problems such as delayed response and subjective judgment, and is difficult to meet the requirements of fairness and accuracy. Especially in large-scale sports tests, manual measurement is inefficient and prone to errors caused by fatigue.

[0003] Although high-speed video technology can record continuous images of the long jump process, how to automatically identify the landing key frames from a large number of video frames is still a challenging task. This is mainly because the posture changes of students during the jump are complex and diverse, and there are significant individual differences. For example, some students may use a landing method of bending their knees to cushion the landing, while others may stay upright and land. This diversity makes it difficult to determine the landing moment only by simple posture features.

[0004] In addition, the complexity of the actual test environment further increases the difficulty of landing recognition. Differences in indoor and outdoor site conditions, changes in light intensity, weather factors (such as wind during outdoor testing), etc. may all affect image quality and motion trajectory. These environmental factors not only affect the clarity and contrast of the image, but may also change the student's jumping performance, thereby affecting the recognition of landing features.

[0005] Some current computer vision methods, such as posture estimation based on deep learning, perform well in static image analysis, but may suffer from insufficient accuracy or poor real-time performance in high-speed motion scenes. Traditional motion detection algorithms, although computationally efficient, often have difficulty coping with complex backgrounds and rapidly changing lighting conditions. Summary of the invention

[0006] Based on this, it is necessary to provide an automatic recognition method for the landing moment of standing long jump based on optical flow analysis that can cope with complex backgrounds and rapidly changing lighting conditions to address the above technical problems.

[0007] A method for automatically identifying the landing moment of standing long jump based on optical flow analysis comprises the following steps:

[0008] A high-speed camera is used to capture a high-frame rate video sequence of the student during the entire jumping process; the high-speed camera is arranged in the student's long jump landing area according to preset rules;

[0009] Preprocessing the high frame rate video sequences in sequence to obtain high frame rate video optimized sequences;

[0010] Based on the high frame rate video optimization sequence, a pixel-level dynamic optical flow field sequence between adjacent frames of the high frame rate video optimization sequence is calculated;

[0011] Recursively decompose each optical flow field in the pixel-level dynamic optical flow field sequence to obtain a series of motion components at different motion scales;

[0012] Calculate the complexity index time series corresponding to the motion component;

[0013] Based on the complexity index time series, identify and extract the key motion features that characterize the student's landing moment;

[0014] Based on a multi-level heuristic decision tree model, key motion features are processed to preliminarily identify potential landing moment frames;

[0015] Collect environmental factors and build an environmental compensation model based on environmental factors and pixel-level dynamic optical flow field sequences;

[0016] Based on the environmental compensation model, potential landing instant frames and pixel-level dynamic optical flow field sequences are processed to obtain landing key frames, and the students' long jump scores are calculated based on the landing key frames.

[0017] In one embodiment, based on the high frame rate video optimization sequence, the step of calculating the pixel-level dynamic optical flow field sequence between adjacent frames of the high frame rate video optimization sequence includes the steps of:

[0018] For each pair of adjacent frames in the high frame rate video optimization sequence, a 5-layer pyramid is constructed; the resolution of the next layer in the 5-layer pyramid is half of the resolution of the previous layer;

[0019] The five-layer pyramid is optimized three times from top to bottom to obtain a pixel-level dynamic optical flow field sequence;

[0020] Pixel-level dynamic optical flow field sequences are optimized using mirror filling technology and forward-backward consistency check.

[0021] In one embodiment, the step of recursively decomposing each optical flow field in the pixel-level dynamic optical flow field sequence to obtain a series of motion components at different motion scales includes the steps of:

[0022] Based on the pixel-level dynamic optical flow field sequence, a Gaussian pyramid is constructed; the upper layer in the Gaussian pyramid is the result of the next low-pass filtering and downsampling;

[0023] Calculate the difference between adjacent layers of the Gaussian pyramid to construct a Laplacian pyramid; the Laplacian pyramid represents detailed information at a specific scale;

[0024] Decompose the top layer of the Gaussian pyramid and all layers of the Laplacian pyramid to obtain motion components at a series of different motion scales.

[0025] In one embodiment, the step of calculating the complexity index time series corresponding to the motion component includes the steps of:

[0026] Calculate the horizontal direction sample entropy time series and the vertical direction sample entropy time series of the motion component;

[0027] Calculate the time series of complexity ratios between different motion scales corresponding to motion components;

[0028] Based on the horizontal sample entropy time series, the vertical sample entropy time series and the complexity ratio time series, the complexity index time series is obtained.

[0029] In one embodiment, the step of identifying and extracting key motion features characterizing the moment of a student landing based on the complexity index time series includes the steps of:

[0030] Analyze the complexity index time series to obtain the characteristic information corresponding to different motion scales;

[0031] Based on the characteristic information, design the first comprehensive indicator sequence;

[0032] A peak detection algorithm is used, and based on time interval constraints, complexity pattern constraints, and continuity constraints, a local maximum value is found in the first comprehensive indicator sequence as the landing instant frame;

[0033] The key feature vectors are extracted from the landing moment frame, and the key feature vectors are normalized to obtain the key motion features that characterize the moment when the students land.

[0034] In one embodiment, the step of processing key motion features based on a multi-level heuristic decision tree model to preliminarily identify potential landing moment frames includes the following steps:

[0035] Constructing a multi-level heuristic decision tree model, which is used to process the extracted key motion features;

[0036] In the first level, comprehensive indicators in key motion features are evaluated, and moments above the preset threshold are marked as preliminary potential landing moments;

[0037] In the second level, for the initial potential landing moment, the complexity change rate of key motion features at different scales is analyzed to further screen out the secondary potential landing points;

[0038] In the third level, for the secondary potential landing points, the changes in the complexity ratios between different scales in the key motion features are evaluated to obtain the high-level potential landing points;

[0039] In the fourth level, for high-level potential landing points, the dynamic time warping algorithm is used to calculate the similarity between their timing patterns and the predefined typical landing patterns to determine the initial candidate landing moments; global time constraints are applied to the initial candidate landing moments to ensure that there is a sufficient time interval between adjacent initial candidate landing moments to obtain the candidate landing moments; the support vector machine classifier is used to verify the candidate landing moments to further improve the recognition accuracy; the time point sequence of the candidate landing moments that have been screened and verified at all levels and the confidence scores corresponding to the candidate landing moments are output.

[0040] In one embodiment, the step of collecting environmental factors and constructing an environmental compensation model based on the environmental factors and a pixel-level dynamic optical flow field sequence includes the following steps:

[0041] Standardize environmental factors;

[0042] Calculate the Pearson correlation coefficient between the standardized environmental factors and the pixel-level dynamic optical flow field sequence;

[0043] Perform principal component analysis on standardized environmental factors whose Pearson correlation coefficient is greater than a preset value to obtain environmental principal components;

[0044] A multivariate regression model was established with the principal components of the environment as the independent variable and the pixel-level dynamic optical flow field sequence as the dependent variable.

[0045] Sensitivity analysis was performed by calculating the partial derivatives of standardized environmental factors whose Pearson correlation coefficients were greater than the preset values, to quantify the influence of standardized environmental factors whose Pearson correlation coefficients were greater than the preset values;

[0046] Based on the multiple regression model and impact degree, an environmental compensation model is constructed.

[0047] In one embodiment, the construction of the environmental compensation model further includes the following steps:

[0048] The root mean square error of the multivariate regression model was evaluated, and if it exceeded a preset threshold, a nonlinear model was adopted;

[0049] Use nonlinear models to refit the relationship between the principal components of the environment and the pixel-level dynamic optical flow field sequence;

[0050] Sensitivity analysis is performed by calculating the partial derivatives of the principal components of the environment with respect to the pixel-level dynamic optical flow field sequence;

[0051] Based on the sensitivity analysis results, the environmental deviation estimation function is constructed; the time weighting strategy is introduced to construct the time weighting function;

[0052] Applying the time weight function to the environmental deviation estimation function, a time-varying environmental compensation model is obtained;

[0053] A time-varying environment compensation model is used to correct the pixel-level dynamic optical flow field sequence.

[0054] In one embodiment, the steps of processing potential landing instant frames and pixel-level dynamic optical flow field sequences based on the environment compensation model, obtaining landing key frames, and calculating the student's long jump score based on the landing key frames include:

[0055] Within the preset time window, the pixel-level dynamic optical flow field sequence is corrected based on the environmental compensation model;

[0056] Based on the corrected pixel-level dynamic optical flow field sequence, a multi-scale analysis method is used to locate the potential landing moment frame;

[0057] A peak detection algorithm is used to search for the local maximum value of the second comprehensive indicator around the positioning result, and the local maximum value is used as the landing key frame;

[0058] Convert the two-dimensional coordinates of the landing keyframe into the three-dimensional space coordinates of the landing;

[0059] Calculate the student's long jump score based on the three-dimensional spatial coordinates of landing and take-off.

[0060] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and is characterized in that the steps of the above method are implemented when the processor executes the computer program.

[0061] One of the above technical solutions has the following advantages and beneficial effects:

[0062] The present invention relates to an automatic recognition method for the landing moment of a standing long jump based on optical flow analysis, which comprises the following steps: using a high-speed camera to collect a high-frame rate video sequence of a student during the entire jumping process; arranging the high-speed camera in the long jump landing area of ​​the student according to preset rules; sequentially performing distortion correction processing, noise suppression processing, contrast optimization processing, white balance processing and gamma correction processing on the high-frame rate video sequence to obtain a high-frame rate video optimization sequence; based on the high-frame rate video optimization sequence, calculating a pixel-level dynamic optical flow field sequence between adjacent frames of the high-frame rate video optimization sequence; recursively decomposing each optical flow field in the pixel-level dynamic optical flow field sequence to obtain a series of different motion The method comprises the following steps: first, collecting motion components of different scales; calculating the complexity index time series corresponding to the motion components; identifying and extracting the key motion features characterizing the student's landing moment based on the complexity index time series; processing the key motion features based on a multi-level heuristic decision tree model to preliminarily identify potential landing moment frames; collecting environmental factors, and constructing an environmental compensation model based on the environmental factors and pixel-level dynamic optical flow field sequences; processing potential landing moment frames and pixel-level dynamic optical flow field sequences based on the environmental compensation model to obtain landing key frames, and calculating the student's long jump results based on the landing key frames, thereby achieving efficient automation throughout the process, comprehensively and meticulously analyzing the movements, and achieving accurate and objective measurement of results. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a flowchart of a method for automatically identifying the landing moment of a standing long jump based on optical flow analysis in an embodiment of the present application.

[0064] Figure 2 Schematic diagram of the internal structure of a computer device in an embodiment of the present application. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0066] In one embodiment, Figure 1 As shown, a method for automatically identifying the landing moment of standing long jump based on optical flow analysis is provided, comprising the following steps:

[0067] Step S110, using a high-speed camera to capture a high-frame rate video sequence of the student during the entire jumping process; the high-speed camera is arranged in the student's long jump landing area according to a preset rule.

[0068] In the standing long jump test of school physical education classes, in order to accurately capture the complete jumping process of students, especially the critical landing moment, a high-performance high-speed camera system needs to be deployed in the long jump landing area. The core of the system is a high-speed camera that can achieve a collection speed of at least 1000 frames per second at full HD resolution, fully meeting the needs of capturing fast-moving details.

[0069] Among them, the preset rules are that the height of the center of the high-speed camera lens from the ground is between 0.3 meters and 0.5 meters, the optical axis of the high-speed camera is perpendicular to the center line of the long jump area and the distance is between 5 meters and 8 meters; the focal length of the high-speed camera is configured to be between 24 mm and 50 mm, the aperture of the lens is configured to be between f / 2.8 and f / 5.6, and the shutter speed is configured to be between 1 / 2000 seconds and 1 / 4000 seconds.

[0070] Specifically, the placement of the high-speed camera is crucial. It should be mounted on a tripod, with the center of the lens parallel to the ground and the height set between 0.3m and 0.5m, which is close to the expected landing height of most students. In addition, the optical axis of the camera lens is perpendicular to the center line of the long jump area, with a horizontal distance of 5m to 8m. This setting ensures that the complete jumping process of the students from the shortest to the tallest is captured.

[0071] For the best image quality, a prime lens with a focal length of 24mm to 50mm can be used. It is worth noting that this focal length range provides a field of view of about 3 meters to 5 meters at a given shooting distance, which is enough to cover the entire area from the take-off point to the farthest possible landing point. In addition, the aperture setting of the lens should be between f / 2.8 and f / 5.6, which allows enough light to enter to maintain a low ISO value (sensitivity value) while ensuring sufficient depth of field.

[0072] Shutter speed is a key parameter to ensure that fast movements are captured clearly. Considering that the maximum speed of a high school student in the standing long jump may reach 6 to 8 meters per second, the shutter speed should be set between 1 / 2000 and 1 / 4000 seconds. This setting can effectively freeze the action and avoid motion blur while retaining sufficient exposure.

[0073] In addition, in order to obtain good exposure under high-speed photography conditions, lighting is a factor that cannot be ignored. Two sets of high-frequency LED (Light Emitting Diode) lamps are arranged on both sides of the long jump area. The power of each lamp should be between 500W (watts) and 1000W, and the color temperature should be 5600K, which matches the color of daylight. The frequency of these lamps should reach more than 1000Hz (hertz), which is far higher than the frame rate of high-speed cameras, so there will be no flickering effect in the video. The angle of the lights should be carefully calculated to minimize the shadows on the students and avoid glare that may interfere with the students.

[0074] The high frame rate video sequence includes a 3-second to 5-second high frame rate video sequence of the student's preparation for jumping saved in the buffer before the high-speed camera starts recording, and a 5-second to 7-second high frame rate video sequence of the student's jumping after the high-speed camera starts recording. It should be noted that in actual shooting, the pre-trigger recording mode is adopted. Specifically, the high-speed camera will continue to save the most recent 3 to 5 seconds of images in the buffer. When the operator triggers the recording through the remote control, the high-speed camera will save the images 3 to 5 seconds before the trigger and 5 to 7 seconds after the trigger, totaling 8 to 12 seconds of high-speed video sequence. This method ensures that no key actions will be missed, while also avoiding the generation of too much redundant data.

[0075] In one example, before the step of using a high-speed camera to capture a high-frame rate video sequence of the student's entire jumping process, the step includes: spatially calibrating and color calibrating the high-speed camera.

[0076] Before each formal test, a series of calibration and testing procedures are carried out. The first is spatial calibration, placing a 1-meter-long standard ruler in the shooting area and shooting a calibration picture. The calibration picture will be used for subsequent image analysis to ensure the accuracy of the measurement results. Next is color calibration, using a standard 24-color card for white balance adjustment and color correction to ensure that the captured picture can truly reflect the color of the scene. In order to cope with possible light changes (especially in outdoor tests), an automatic aperture control system is installed on the camera. The automatic aperture control system can automatically adjust the aperture size within 1 / 60 second according to the changes in ambient light to maintain the exposure consistency of the picture. The automatic aperture control system is used in conjunction with the automatic gain control (AGC) of the high-speed camera to largely offset the impact of light fluctuations.

[0077] After the step of using a high-speed camera to capture a high-frame rate video sequence of the student's entire jumping process, the method includes the following steps: after acquiring the high-frame rate video sequence, the high-frame rate video sequence is named according to the test date, student number and jumping order, and saved. It should be noted that in terms of data management, an automated file naming and organization system has been developed. Each video file is automatically named according to the test date, student number and jumping order. Through the above high-speed camera system, a high-frame rate video sequence of the student's complete jumping process was successfully captured.

[0078] Step S120, preprocessing the high frame rate video sequence in sequence to obtain a high frame rate video optimized sequence.

[0079] Preprocessing includes distortion correction, noise suppression, contrast optimization, white balance and gamma correction.

[0080] The distortion correction process is performed based on the following steps:

[0081] A 9×6 checkerboard pattern was used and 20 images were taken at different angles and distances using a high-speed camera.

[0082] The calibrateCamera function in the OpenCV library is used to analyze the position deviation of the checkerboard corner points in the image and obtain the distortion coefficient of the high-speed camera.

[0083] Based on the distortion coefficient, the cv2.undistort function is applied to each frame image in the high frame rate video sequence for distortion correction.

[0084] Specifically, due to the optical characteristics of the high-speed camera lens, a certain degree of distortion is inevitably present in high-frame-rate video sequences, mainly radial distortion and tangential distortion. In order to correct these distortions, the high-speed camera is first calibrated. The calibration process uses a standard 9×6 checkerboard pattern and takes 20 images at different angles and distances. The position deviation of the checkerboard corner points of the image in the high-frame-rate video sequence is analyzed by the calibrateCamera function in the OpenCV library to obtain the intrinsic parameter matrix and distortion coefficient of the lens. After obtaining the distortion coefficient, the cv2.undistort function is applied to each frame of the high-frame-rate video sequence for distortion correction. This process is based on a polynomial model and remaps the position of each pixel in the image of the high-frame-rate video sequence to eliminate the deformation caused by the lens and ensure the accuracy of the image geometry.

[0085] The noise suppression process is performed based on the following steps:

[0086] The high frame rate video sequence with distortion correction is processed by Gaussian filter, bilateral filter and median filter in sequence.

[0087] Specifically, during the high-speed video recording process, due to the short exposure time, more Gaussian noise and a small amount of salt and pepper noise were introduced. In order to effectively suppress these noises while retaining the detailed information of the image, a two-stage filtering strategy was adopted. First, a 5×5 kernel Gaussian filter was used to preliminarily smooth the image, and the standard deviation of the filter was set to 1.5. This step can effectively reduce Gaussian noise, but may cause slight blurring of the image. To compensate for this, an edge-preserving filter, the bilateral filter, was then applied. The spatial domain standard deviation of the bilateral filter was set to 75, and the value domain standard deviation was set to 75. This filter takes into account the spatial distance and gray value differences of pixels, and keeps the edges sharp while further smoothing the image. In particular, a 3×3 median filter was used for processing the residual salt and pepper noise. This combined filtering strategy can effectively suppress noise while retaining the detailed information of the image to the maximum extent, ensuring that the contours of the students' movements are clearly discernible.

[0088] The contrast optimization process is performed based on the following steps.

[0089] The high frame rate video sequence after noise suppression is divided into 8×8 grids.

[0090] Perform histogram equalization on the grid individually.

[0091] The result of histogram equalization by merging adjacent grids using bilinear interpolation.

[0092] Specifically, the contrast-limited adaptive histogram equalization (CLAHE) algorithm is used to enhance the image contrast. The advantage of the CLAHE algorithm is that it can locally enhance the image contrast and avoid the over-enhancement problem that may be caused by global histogram equalization. In the specific implementation, the image is divided into 8×8 grids, and the size of each grid is about 240×135 pixels (based on a resolution of 1920×1080). Histogram equalization is performed on each grid separately, and the contrast limit threshold is set to 3.0 to prevent the noise from being over-amplified. Then the bilinear interpolation method is used to merge the results of adjacent grids to eliminate the artificial boundaries that may be generated by the block processing. This local adaptive method can effectively enhance the local contrast of the image, making the outline and details of the students more clearly discernible, especially when the students are wearing sportswear with a color similar to the background.

[0093] White balance is performed based on the following steps:

[0094] The average values ​​of the R channel, G channel and B channel of the high frame rate video sequence after contrast optimization processing are calculated respectively.

[0095] The ratio of the average value to the ideal gray value is used as the gain coefficient corresponding to the R channel, the G channel and the B channel respectively.

[0096] The gain factor is applied to each pixel of the contrast-optimized high frame rate video sequence.

[0097] Specifically, in order to ensure the color accuracy of the image, automatic white balance processing is performed. An improved gray world hypothesis algorithm is used, assuming that the average values ​​of the three channels R, G, and B in the image should be close to gray. First, the average values ​​of the R, G, and B channels of the entire image are calculated, and then the ratios of these average values ​​to the ideal grayscale value (usually 128) are calculated as the gain coefficients of each channel. To avoid overcorrection in extreme cases, the gain coefficients are limited to between 0.5 and 2.0. Finally, these gain coefficients are applied to each pixel of the original image to achieve overall color balance. This step ensures that videos shot under different lighting conditions (such as cloudy days or strong sunlight) can maintain consistent color performance, which helps the stability of the subsequent analysis algorithm.

[0098] Gamma correction is performed based on the following steps:

[0099] The gamma value of the high frame rate video sequence after white balance processing is adjusted by using a table lookup method to obtain a high frame rate video optimized sequence.

[0100] Specifically, the Gamma value is adjusted dynamically based on the average brightness of the image. When the image is dark overall, a Gamma value less than 1 (such as 0.8) is used to brighten the dark details; when the image is bright overall, a Gamma value greater than 1 (such as 1.2) is used to enhance the contrast. Gamma correction is implemented using a lookup table (LUT) method to pre-calculate the pixel mapping relationship under different Gamma values ​​to increase processing speed. It is worth mentioning that this step is particularly important for tests that adapt to different time periods (such as early morning or evening) to ensure that students' movements can be clearly captured under various lighting conditions.

[0101] Through the above series of sophisticated image preprocessing operations, the image quality in high frame rate video sequences is significantly improved.

[0102] Step S130, based on the high frame rate video optimization sequence, calculating a pixel-level dynamic optical flow field sequence between adjacent frames of the high frame rate video optimization sequence.

[0103] The dynamic optical flow field is a dense vector field that describes the motion of objects in an image. It can accurately capture the subtle changes in students' movements during the jumping process, including the motion characteristics of key stages such as take-off, mid-air posture adjustment, and landing.

[0104] It should be pointed out that the optical flow calculation is based on three basic assumptions: constant pixel brightness, small object motion, and spatial coherence. The assumption of constant pixel brightness means that the brightness of the same object in adjacent frames should remain unchanged, which allows motion to be estimated by tracking brightness patterns. The assumption of small object motion means that the displacement of objects between adjacent frames will not be too large, which allows searching for corresponding points in local areas. The assumption of spatial coherence indicates that adjacent pixels tend to have similar motion, which helps to make more robust estimates in the presence of noisy or blurred areas.

[0105] Considering the characteristics of the standing long jump action, especially the rapid motion that may be included therein, the Fanabac dense optical flow algorithm is preferably used. This algorithm is based on polynomial expansion and can effectively handle a large range of motion, and is very suitable for capturing rapid motion changes during the jump process. The core idea of ​​the Fanabac algorithm is to represent the signal in each pixel area with a quadratic polynomial: f(x)=x^TAx+b^Tx+c, where x is the coordinate vector, A is a symmetric matrix, b is a vector, and c is a scalar. By comparing the polynomial coefficients of the corresponding areas in the two frames of images, the displacement of the pixels can be accurately estimated.

[0106] The specific steps of adopting the Fanabac dense optical flow algorithm are to calculate the pixel-level dynamic optical flow field sequence between adjacent frames of the high frame rate video optimization sequence based on the high frame rate video optimization sequence, including the steps of: constructing a 5-layer pyramid for each pair of adjacent frames in the high frame rate video optimization sequence; the resolution of the next layer in the 5-layer pyramid is half of the resolution of the previous layer; the 5-layer pyramid is iteratively optimized 3 times from top to bottom to obtain a pixel-level dynamic optical flow field sequence; and the pixel-level dynamic optical flow field sequence is optimized by using the mirror filling technology and the forward-backward consistency check.

[0107] In order to optimize the performance of the algorithm, the following key parameters are used: the number of pyramid layers is set to 5 to allow the algorithm to handle a wide range of motion; the pyramid scaling factor is 0.5, and the resolution of each layer of the image is half of the previous layer; the number of iterations is set to 3 to strike a balance between accuracy and computational efficiency; the window size is 21 pixels to provide sufficient local information; the polynomial expansion neighborhood size is 7 to ensure accurate local signal description; the Gaussian filter standard deviation is set to 1.5 to effectively suppress noise while retaining important structural information. The selection of these parameters ensures that the algorithm can obtain high-quality optical flow estimation when dealing with complex motions such as standing long jump.

[0108] Specifically, the optical flow calculation process starts with building an image pyramid. For each pair of adjacent frames of the input, a 5-layer pyramid is constructed, and the resolution of each layer is half of the previous layer. This multi-scale approach not only helps to handle a wide range of motion, but also captures motion details at different scales. Starting from the top layer of the pyramid, the initial flow field is calculated for each pixel. This process uses a 7×7 neighborhood for polynomial expansion to obtain coefficients A_1, b_1 and A_2, b_2. Then, the displacement d of the pixel is calculated by solving the equation d=-(A_1+A_2)^(-1)*(b_2-b_1).

[0109] In each layer of the pyramid, iterative optimization is performed to improve the accuracy of the flow field estimation. In each iteration, the previous estimate is used to compensate the second frame image, and then the displacement is recalculated. This process is repeated 3 times to ensure convergence to a stable solution. After the calculation is completed, the obtained flow field is upsampled to the next layer of the pyramid as the initial estimate of that layer. This process is repeated until the bottom layer of the pyramid is processed. The final flow field is the desired pixel-level optical flow field, which is denoted as F(t, t+1), representing the optical flow field from frame t to frame t+1.

[0110] It is worth noting that the calculated optical flow field F(t, t+1) is a two-dimensional matrix with the same size as the input image, usually 1920×1080 (full HD resolution). Each element in the matrix is ​​a two-dimensional vector (u, v), representing the displacement of the corresponding pixel in the x and y directions. This representation method intuitively describes the movement of each point in the image.

[0111] A number of quality control measures were implemented to ensure the accuracy and reliability of the optical flow calculations. A mirror filling technique was used at the edges of the image to ensure that edge pixels also received accurate optical flow estimates. This approach can avoid inaccurate motion estimates at the edges of the image, especially in cases where the hands and feet of athletes may be close to the edge of the image. Statistical methods were used to detect and process possible outliers to prevent them from affecting subsequent analysis. Outliers may arise from image noise, occlusion, or rapid motion. By identifying and processing these outliers, the quality of the overall optical flow field can be improved.

[0112] In addition, a forward-backward consistency check is performed, that is, the optical flow from the first frame to the second frame is calculated, and then the optical flow from the second frame back to the first frame is calculated, and the consistency of the two calculation results is compared. If they are inconsistent, these areas will be marked and specially processed to improve the reliability of the overall estimate. This two-way check can effectively identify unreliable estimates caused by occlusion or fast motion, thereby improving the overall quality of the optical flow field.

[0113] Finally, the above process is repeated for each pair of adjacent frames in the high frame rate video optimization sequence to obtain a series of F(t, t+1), which together constitute the pixel-level dynamic optical flow field sequence F_original(t). Each F_original(t) is associated with the corresponding frame index and precise timestamp, which ensures the temporal continuity and traceability of the data. This sequence records in detail the motion state of the student at every moment during the jump, including key information such as speed and direction.

[0114] Step S140 , recursively decompose each optical flow field in the pixel-level dynamic optical flow field sequence to obtain a series of motion components of different motion scales.

[0115] For example, a Gaussian-Laplacian pyramid structure is used for recursive decomposition, which can effectively decompose motion information into different spatial scales. The Gaussian-Laplacian pyramid is a multi-scale representation method that is widely used in computer vision and image processing, and is particularly suitable for analyzing complex scenes containing multi-scale motions, such as overall displacement and local limb movements in standing long jump.

[0116] In one example, recursively decomposing each optical flow field in a pixel-level dynamic optical flow field sequence to obtain a series of motion components of different motion scales includes the following steps: constructing a Gaussian pyramid based on the pixel-level dynamic optical flow field sequence; the previous layer in the Gaussian pyramid is the result of the next low-pass filtering and downsampling; calculating the difference between adjacent layers of the Gaussian pyramid to construct a Laplacian pyramid; the Laplacian pyramid represents detail information of a specific scale; decomposing the top layer of the Gaussian pyramid and all layers of the Laplacian pyramid to obtain a series of motion components of different motion scales.

[0117] Specifically, a Gaussian pyramid is constructed starting from a high frame rate video optimization sequence. Each layer of the Gaussian pyramid is the result of low-pass filtering and downsampling of the next layer, and this process can be expressed as: F_i=downsample(gaussian_blur(F_(i-1))). Where i represents the pyramid level, ranging from 0 to N-1, and N is the total number of layers of the pyramid. Preferably, N=5 is selected, that is, a 5-layer pyramid is constructed. This number of layers is selected based on the analysis of the complexity of the fixed long jump action, and can effectively capture various scales of motion from the overall jump trajectory to tiny limb adjustments. The selection of a 5-layer pyramid is a trade-off between computational complexity and information retention, which can fully decompose motion information without introducing too much computational burden.

[0118] Specifically, the Gaussian pyramid construction process is as follows: First, F_0=F_original(t), maintaining the full resolution of the pixel-level dynamic optical flow field sequence. Then, F_1=downsample(gaussian_blur(F_0)), this is the first downsampling, and the resolution is halved. Next, F_2=downsample(gaussian_blur(F_1)), this is the second downsampling, and the resolution is 1 / 4 of the original. Again, F_3=downsample(gaussian_blur(F_2)), this is the third downsampling, and the resolution is 1 / 8 of the original. Finally, F_4=downsample(gaussian_blur(F_3)), this is the fourth downsampling, and the resolution is 1 / 16 of the original. This layer-by-layer downsampling process allows each layer to contain motion information of different scales, and everything from the subtlest local motion to the most macroscopic overall displacement is effectively captured and represented.

[0119] It should be pointed out that before each downsampling, Gaussian blur is first applied to suppress high-frequency information and prevent aliasing effects. Gaussian blur uses a 5×5 kernel with a standard deviation of σ=1.0. The selection of this kernel size and standard deviation is based on a large number of experimental results, which can achieve a good balance between retaining sufficient details and effectively suppressing noise. The calculation formula of the Gaussian kernel is: G(x, y)=(1 / (2πσ²))*exp(-(x²+y²) / (2σ²)), where x and y are the coordinates of the pixel relative to the center of the kernel. This Gaussian filter can not only effectively smooth the image and reduce the impact of noise, but also maintain the main structural features of the image, laying the foundation for subsequent multi-scale analysis.

[0120] During the downsampling process, mean pooling with a factor of 2 is used, that is, the average value of each 2×2 pixel block is taken as the corresponding single pixel value of the next layer. This method can better preserve the overall characteristics of the image than simple point sampling, because it considers the information of all pixels in the local area instead of simply discarding some pixels. Mean pooling can maintain the main structure and texture information of the image while reducing the resolution, which is crucial to maintaining the continuity and consistency of motion information.

[0121] After constructing the Gaussian pyramid, the next step is to construct the Laplacian pyramid. The Laplacian pyramid is constructed by calculating the difference between adjacent layers of the Gaussian pyramid, and each layer represents detailed information at a specific scale. The construction process can be expressed as: L_i=F_i-upsample(F_(i+1)). The upsample operation uses bilinear interpolation to enlarge F_(i+1) to the same size as F_i. Bilinear interpolation can provide smooth enlargement results while ensuring computational efficiency, avoiding the jagged edges that may be caused by simply copying pixels.

[0122] Specifically, the Laplacian pyramid calculation process is: first, L_3=F_3-upsample(F_4), this layer contains larger-scale motion details. Then, L_2=F_2-upsample(F_3), this layer contains medium-scale motion information. Next, L_1=F_1-upsample(F_2), this layer contains smaller-scale local motion. Finally, L_0=F_0-upsample(F_1), this layer contains the most subtle local motion details. Each layer Li contains motion details at that scale, which are smoothed out at the coarser scale F_(i+1). In this way, the Laplacian pyramid effectively separates motion information of different scales, so that subsequent analysis can be performed on motion features of specific scales.

[0123] The final decomposition result includes the layers L_0, L_1, L_2, L_3 of the Laplacian pyramid, and the top layer F_4 of the Gaussian pyramid, which together constitute the complete decomposition of the original optical flow field. The physical meaning and characteristics of each layer are as follows: The L_0 layer represents the most subtle local movement, corresponding to the slight shaking of the student's limbs, and has the smallest spatial range. This layer can capture details such as the slight tremor of the fingers or the subtle adjustment of the toes. These tiny movements may reflect the balance adjustment of the athlete when taking off or landing. The L_1 layer represents smaller-scale local movements, such as the swing of the arms and legs. This layer can reflect the swing amplitude of the arms when taking off or the cushioning action of the legs when landing. This information is crucial for evaluating the quality of the athlete's technical movements. The L_2 layer represents medium-scale movements, such as the twisting of the torso. This layer can capture the coordination of the body during the jump, such as the coordinated movements of the upper and lower body during the take-off, which is very valuable for evaluating the overall coordination and technical level of the athlete. The L_3 layer represents larger-scale movements, such as the overall take-off action. This layer reflects the main trajectory of the jump, including key information such as the angle of take-off and initial speed, which are directly related to the distance and quality of the long jump. The F_4 layer represents the overall movement of the largest scale, reflecting the overall displacement of the student in the picture, with the largest spatial range. This layer can be used to estimate the overall distance and trajectory of the jump, providing an intuitive representation of the long jump performance.

[0124] Step S150, calculating the complexity index time series corresponding to the motion component.

[0125] For example, using sample entropy (SampEn) as the main complexity indicator effectively quantifies the irregularity of the time series and is suitable for describing the complexity of motion. Sample entropy is a statistic used to evaluate the complexity or randomness of a time series. It is based on the concept that in a complex system, similar patterns are unlikely to recur. Therefore, the higher the sample entropy value, the more complex or random the time series is; the lower the value, the more regular or predictable the time series is.

[0126] In one example, the step of calculating a complexity index time series corresponding to a motion component includes the steps of: calculating a horizontal sample entropy time series and a vertical sample entropy time series of the motion component; calculating a complexity ratio time series between different motion scales corresponding to the motion component; and obtaining a complexity index time series based on the horizontal sample entropy time series, the vertical sample entropy time series and the complexity ratio time series.

[0127] For each scale of motion component L_i (i=0, 1, 2, 3) and F_4 obtained in step S140, the sample entropy time series in the horizontal direction (u) and vertical direction (v) are calculated respectively:

[0128] SampEn_u(L_i, t) = Sample entropy time series in the horizontal direction

[0129] SampEn_v(L_i, t) = Sample entropy time series in the vertical direction

[0130] Where t represents the time point, corresponding to the center frame of each sliding window. This approach of calculating horizontal and vertical directions separately allows capturing the complexity of motion in different directions.

[0131] In order to comprehensively consider the complexity in both directions, the overall complexity index time series is further defined:

[0132] C(L_i,t)=√(SampEn_u(L_i,t)²+SampEn_v(L_i,t)²)

[0133] This formula actually calculates the Euclidean norm of the sample entropy time series in the horizontal direction and the sample entropy time series in the vertical direction. C(L_i, t) reflects the overall complexity of the motion component of the i-th scale at time t. It is worth noting that the larger the value, the more complex or irregular the motion at that moment. This method of considering both directions can more comprehensively describe the complexity of the motion, because the motion in the standing long jump is three-dimensional, and the motion in the horizontal and vertical directions may have different complexity characteristics.

[0134] Through this method, we finally get the complexity index time series of the motion component at each scale:

[0135] {C(L_i, 1), C(L_i, 1+S), C(L_i, 1+2S),..., C(L_i, 1+(TW)}

[0136] These time series not only reflect the overall complexity of the motion components at each scale, but also show the temporal variation pattern of the complexity. For example, it may be observed that the complexity of certain scales has significant peaks or valleys at the moment of take-off and landing.

[0137] In addition, to further enhance the reliability of the analysis, the time series of complexity ratios between scales are calculated. R(i, j, t) = C(L_i, t) / C(L_j, t) represents the ratio of the complexity of the i-th scale to the j-th scale at time t. These ratio series provide information about the changes in the relative complexity between different scales of motion, which helps to identify the transition points of motion states, such as the moment of take-off and landing. For example, it may be found that at certain critical moments, the complexity of small-scale motion (such as L_0 or L_1) changes significantly relative to the complexity of large-scale motion (such as L_3 or F_4).

[0138] Get the sample entropy time series based on the following steps:

[0139] First, set the sliding window size W (50 frames, about 0.1 seconds, assuming the video frame rate is 500fps) and step size S (10 frames). The selection of the window size W is based on the analysis of the characteristics of the fixed long jump action to ensure that sufficient motion details can be captured without over-smoothing the temporal changes. The selection of the step size S strikes a balance between computational efficiency and temporal resolution. For a video sequence with a total number of frames T, the sliding window starts from the first frame and moves S frames each time until the end of the window reaches T frames.

[0140] Perform the following detailed steps for the data within each window:

[0141] 1. In the current window, for a given time series {x_i}, i=1, 2, ..., W, select an embedding dimension m and a similarity tolerance r. Preferably, set m=2 and r=0.2*σ, where σ is the standard deviation of the time series in the window. The embedding dimension m determines the length of the compared patterns, while the similarity tolerance r defines the threshold at which two patterns are considered similar. The choice of m=2 allows capturing short-term dependencies in the sequence, while the setting of r=0.2*σ allows the similarity judgment to adapt to local fluctuations in the data.

[0142] 2. Next, construct the template vector: X_i(m)={x_i, x_(i+1), ..., x_(i+m-1)}

[0143] For example, for the time series {1.2, 3.4, 2.5, 4.6, 5.2, 3.1}, when m=2, the following template vector is obtained:

[0144] X_1(2)={1.2,3.4}

[0145] X_2(2)={3.4, 2.5}

[0146] X_3(2)={2.5,4.6}

[0147] X_4(2)={4.6,5.2}

[0148] X_5(2)={5.2,3.1}

[0149] These template vectors represent local patterns in the time series.

[0150] 3. Then, calculate the number of template matches:

[0151] B_i(m) = number of matches X_i(m) / (W-m+1)

[0152] If the maximum absolute difference between two template vectors is less than r, they are considered to match. This step is actually looking for similar patterns.

[0153] The specific matching process is as follows:

[0154] a) For each template vector Xi(m), iterate over all other template vectors Xj(m) (j≠i).

[0155] b) Calculate the absolute difference between corresponding elements of X_i(m) and X_j(m).

[0156] c) If the absolute differences of all corresponding elements are less than r, the two template vectors are considered to match.

[0157] d) Count the number of template vectors that match Xi(m) and divide it by (W-m+1) to get B(m).

[0158] 4. Then, calculate the average matching probability:

[0159] B(m)=(1 / (Wm))*ΣB_i(m)

[0160] This step averages the matching probabilities B_i(m) of all template vectors to obtain the overall matching probability under the current embedding dimension m.

[0161] 5. Repeat steps 2-4, replacing m with m+1, to get B(m+1). This step is to compare the matching probability under different pattern lengths. Increasing the pattern length usually reduces the matching probability because longer patterns are more difficult to find a complete match.

[0162] 6. Finally, calculate the sample entropy of the current window:

[0163] SampEn=-ln(B(m+1) / B(m))

[0164] This formula actually compares the matching probability in m+1 dimension and m dimension. If the matching probability decreases significantly after increasing the dimension (that is, B(m+1) is much smaller than B(m)), the sample entropy value will be larger, indicating that the sequence is more complex; conversely, if the matching probability does not change much after increasing the dimension, the sample entropy value will be smaller, indicating that the sequence is more regular.

[0165] 7. Assign the calculated sample entropy value to the central frame of the current window as the complexity index at that moment. In this way, a complexity index sequence that changes over time is obtained.

[0166] Step S160, based on the complexity index time series, identifying and extracting key motion features that characterize the student's landing moment.

[0167] In one example, the steps of identifying and extracting key motion features characterizing the moment a student lands based on a complexity index time series include the following steps: analyzing the complexity index time series to obtain feature information corresponding to different motion scales; designing a first comprehensive index sequence based on the feature information; using a peak detection algorithm, and based on time interval constraints, complexity pattern constraints, and continuity constraints, finding the local maximum value in the first comprehensive index sequence as the landing moment frame; extracting key feature vectors from the landing moment frame, and normalizing the key feature vectors to obtain key motion features characterizing the moment a student lands.

[0168] First, the complexity index time series obtained in step S150 is analyzed in depth. For each scale i (i=0, 1, 2, 3, 4, corresponding to L_0, L_1, L_2, L_3, F_4), there is a complexity index time series {C(L_i, t)} and a complexity ratio time series {R(i, j, t)}. These series reflect the changes in motion complexity over time at different spatial scales.

[0169] Pay special attention to significant changes in these sequences, especially sharp increases or decreases in complexity, because these usually correspond to important transitions in the state of motion. In order to capture these changes more accurately, calculate the first difference of the complexity index:

[0170] ΔC(L_i,t)=C(L_i,t+1)-C(L_i,t)

[0171] This difference sequence can highlight rapid changes in complexity, which often correspond to key motion events. For example, at the moment of landing, a sharp change in complexity at certain scales may be observed.

[0172] During the analysis, it was found that the complexity changes at different scales often exhibit different characteristics. Specifically:

[0173] 1. For small-scale movements (L_0, L_1), the complexity usually increases sharply at the moment of landing. This reflects the rapid fine-tuning and vibration of the limbs at the moment of landing. For example, the moment of foot contact may cause a significant peak in the complexity of the L_0 scale.

[0174] 2. For the medium scale (L_2), the complexity change may reflect the adjustment process of the body posture. Before and after landing, a fluctuation in L_2 complexity may be observed, indicating the process of the body adjusting from the air posture to the landing posture.

[0175] 3. For large-scale motion (L_3, F_4), the complexity may drop significantly at the moment of landing. This corresponds to the process of the overall motion changing from rapid falling to buffering. Especially for the F_4 scale, there may be an obvious complexity valley at the moment of landing.

[0176] In order to more comprehensively describe the relationship between different scales, the complexity ratio time series R(i, j, t) calculated in step S150 is used. These ratios provide important information about the relative changes between different scales of motion. For example, when R(0, 4, t) = C(L_0, t) / C(F_4, t) suddenly increases, it may indicate that the small-scale motion becomes relatively more complex, which often occurs at the moment of landing. Pay special attention to the sudden changes in these ratios because they may mark the key transition points of the motion state. Here R(i, j, t) represents the ratio of the complexity of the i-th scale to the j-th scale at time t. C(L_i, t) represents the complexity index of the i-th scale at time t. 0 and 4 represent the complexity of the smallest scale (L_0) and the largest scale (F_4), respectively. R(0, 4, t) represents the ratio of the complexity of the smallest scale (L_0) to the complexity of the largest scale (F_4) at time t.

[0177] Based on the above analysis, a comprehensive indicator J(t) (the first comprehensive indicator) is designed to quantify the landing characteristics:

[0178] J(t)=α*ΔC(L_0,t)+β*ΔC(L_1,t)+γ*ΔC(L_3,t)+δ*ΔC(F_4,t)+ε*(R(0,4,t)-R(0,4,t-1))

[0179] This indicator comprehensively considers the complexity changes at different scales and the relative changes between scales, where α, β, γ, δ, and ε are weight coefficients. The selection of these coefficients is based on statistical analysis of a large amount of sample data to maximize the recognition effect of J(t) at the moment of landing. Usually, small-scale motion (L_0 and L_1) is given a larger weight because they are most sensitive to small changes at the moment of landing.

[0180] Next, a peak detection algorithm is used to find local maxima in the J(t) sequence. These peak points are regarded as potential landing moments. However, relying solely on the peak value of J(t) may lead to some misjudgments. Therefore, a series of constraints are introduced to improve the accuracy of recognition:

[0181] 1. Time interval constraint: The time interval between two potential landing points should be greater than the minimum jump period (usually set to about 1 second). This constraint is based on the physiological limitations of normal human jumping movements and helps filter out some false positive detection results.

[0182] 2. Complexity pattern constraint: It requires that the complexity of each scale should meet a specific change pattern before and after landing. For example, it may be required that the complexity of L_0 and L_1 increases significantly at the moment of landing, while the complexity of F_4 decreases significantly. This pattern constraint can effectively distinguish the real landing moment from other actions that may cause complexity changes (such as adjusting posture in the air).

[0183] 3. Continuity constraint: It is required that the small-scale complexity should be kept at a high level within a short period of time after landing (about 0.1-0.2 seconds). This reflects the short buffering and balance adjustment process after landing, which helps to distinguish between real landing and instantaneous actions in the air.

[0184] For each potential landing point t that satisfies these conditions, a key feature vector X(t) is extracted, containing the following raw eigenvalues:

[0185] 1. Comprehensive index J(t)

[0186] 2. Change rate of complexity at each scale:

[0187] ΔC(L_i,t)=C(L_i,t)-C(L_i,t-1),i=0,1,2,3,4

[0188] 3. Complexity ratios and their changes between different scales:

[0189] R(i,j,t)=C(L_i,t) / C(L_j,t),i,j=0,1,2,3,4,i≠j

[0190] ΔR(i,j,t)=R(i,j,t)-R(i,j,t-1)

[0191] 4. For each scale i (i=0, 1, 2, 3, 4, corresponding to L_0, L_1, L_2, L_3, F_4), calculate the average complexity in the short time window centered on time t: C_avg(L_i, t)=(1 / (2w+1))*ΣC(L_i, τ) where the range of τ is [tw, t+w], and w represents the half-width of the window (such as the number of frames corresponding to ±0.1 seconds).

[0192] In order to ensure the comparability of different elements in the feature vector, the extracted original feature set is normalized to obtain the key motion features that characterize the moment when the student lands. For each element x_i in the key feature vector X(t), the following normalization formula is applied:

[0193] x_i_normalized=(x_i-μ_i) / σ_i

[0194] Where x_i is the i-th original eigenvalue in the eigenvector X, μ_i and σ_i are the mean and standard deviation of the feature in all samples, respectively. This normalization process eliminates the scale differences between different features, so that subsequent analysis can fairly consider the contribution of each feature. The normalized feature vector X_normalized contains all the standardized features, each with zero mean and unit variance.

[0195] Step S170, processing key motion features based on a multi-level heuristic decision tree model to preliminarily identify potential landing moment frames.

[0196] In one example, the step of processing key motion features based on a multi-level heuristic decision tree model to preliminarily identify potential landing moment frames includes the following steps:

[0197] A multi-level heuristic decision tree model is constructed, which is used to process the extracted key motion features. In the first level, the comprehensive indicators in the key motion features are evaluated, and the moments above the preset threshold are marked as preliminary potential landing moments. In the second level, for the preliminary potential landing moments, the complexity change rate of different scales in the key motion features is analyzed, and the secondary potential landing points are further screened. In the third level, for the secondary potential landing points, the changes in the complexity ratios between different scales in the key motion features are evaluated to obtain high-level potential landing points. In the fourth level, for the high-level potential landing points, the dynamic time warping algorithm is used to calculate the similarity between their timing patterns and the predefined typical landing patterns to determine the initial candidate landing moments. Global time constraints are applied to the initial candidate landing moments to ensure that there is enough time interval between adjacent initial candidate landing moments to obtain candidate landing moments. The candidate landing moments are verified using a support vector machine classifier to further improve the recognition accuracy. The time point sequence of the candidate landing moments that have been screened and verified at all levels and the confidence scores corresponding to the candidate landing moments are output.

[0198] Specifically, the first level of the multi-level heuristic decision tree model mainly evaluates the comprehensive index J(t) extracted in step S160. This index combines information from multiple aspects, including the complexity of the movement, speed changes, etc., and can therefore better reflect the characteristics of the landing moment. By defining a preset threshold Tj, when J(t)>Tj, t is marked as a preliminary potential landing moment. It is worth noting that this preset threshold Tj can be dynamically adjusted by a data-driven method to adapt to different motion patterns. Preferably, Tj can be set to the 95% quantile of the J(t) sequence, so that only the most significant peak is identified as a potential landing moment. This adaptive threshold setting method can effectively handle the differences in J(t) distribution caused by different athletes and different jumping styles, and improves the applicability and accuracy of the model.

[0199] In the second level of the multi-level heuristic decision tree model, the model deeply analyzes the complexity change rate ΔC(L_i, t) of each scale in the key feature vector extracted in step S160. The analysis at this level is based on the observation that at the moment of landing, the complexity of motion at different scales will show a specific change pattern. In particular, the model focuses on the sharp increase in complexity of small scales (L_0, L_1) and the significant decrease in complexity of large scales (L_3, F_4). The setting of this focus is derived from the understanding of the physical characteristics of the landing action: when landing, the overall movement of the body (large scale) will slow down rapidly, while the local adjustment action (small scale) will become more complex. For this reason, the corresponding thresholds Ts and Tl are defined. When ΔC(L_0, t)>Ts and ΔC(F_4, t)<-Tl, t is further confirmed as a secondary potential landing point. The negative sign "-Tl" here indicates that the complexity of the F_4 scale has dropped significantly, which is consistent with the physical phenomenon that the overall movement speed decreases sharply at the moment of landing. Specifically, the values ​​of Ts and Tl can be determined by analyzing a large amount of sample data, and are usually set to the 75% quantile of the corresponding complexity change. This setting method ensures sensitivity to landing characteristics while avoiding missed detection due to overly strict conditions.

[0200] The third level of the multi-level heuristic decision tree model evaluates the change ΔR(i, j, t) in the complexity ratio between different scales calculated in step S160. The design of this level is based on the assumption that not only the complexity of each scale will change at the moment of landing, but also the complexity ratio between different scales will change significantly. A threshold Tr is defined, and when ΔR(0, 4, t)=R(0, 4, t)-R(0, 4, t-1)>Tr, t is further confirmed as a high-level potential landing point. This level captures the relative changes between movements of different scales and provides additional judgment basis. R(0, 4, t) represents the ratio of the complexity of the smallest scale (L_0) and the largest scale (F_4). Its sudden increase may mean that the local movement becomes more complex relative to the overall movement, which is a typical feature of the moment of landing. Optionally, in addition to R(0, 4, t), the ratio changes of other scale combinations, such as R(1, 3, t) or R(2, 4, t), can also be considered to obtain more comprehensive information. This consideration of multi-scale ratios can improve the model's adaptability to different types of landing patterns.

[0201] In the fourth level of the multi-level heuristic decision tree model, the model uses the average complexity C_avg(L_i, t) in the short time window before and after landing extracted in step S160 to analyze the movement pattern before and after landing. The design purpose of this level is to capture the timing characteristics of the landing process, rather than relying solely on instantaneous indicators. By implementing a dynamic time warping (DTW) algorithm, the similarity between the observed timing pattern and the predefined typical landing pattern is calculated. The advantage of the DTW algorithm is that it can handle nonlinear distortions on the time axis, which is very important for adapting to the changes in landing speeds of different athletes. Only when the similarity exceeds a certain threshold Ts is the moment finally confirmed as the landing moment (initial candidate landing moment). This level can effectively filter out those moments that have passed the inspection of the previous levels but whose overall pattern does not match the typical landing. Preferably, multiple predefined typical landing patterns can be used to adapt to different jumping styles. These typical patterns can be obtained by analyzing a large amount of labeled landing data and can be subdivided according to different athlete groups (such as age, gender, and training level).

[0202] In order to handle possible multiple jumps, the model also introduces a global constraint: the time interval between two identified landing moments should be greater than the minimum jump period Tmin (usually set to 1-2 seconds) to obtain candidate landing moments. This constraint ensures that the model does not misidentify multiple landing points in a single jump. Specifically, the value of Tmin can be adjusted according to the physiological characteristics of students of different age groups. For example, for younger or weaker students, the value of Tmin can be appropriately increased to adapt to their slower jumping frequency.

[0203] In summary, for each moment t, if all of the following conditions are met, t is identified as a landing moment: J(t)>Tj, ΔC(L_0, t)>Ts and ΔC(F_4, t)<-Tl, R(0, 4, t)-R(0, 4, t-1)>Tr, DTW_Similarity(Pattern(t), TypicalLandingPattern)>Ts, and t-tlast_landing>Tmin. This combination of multiple conditions not only improves the accuracy of recognition, but also enhances the robustness of the model, which can adapt to different jumping patterns and environmental conditions.

[0204] In order to further improve the accuracy of the model, a verification step based on machine learning is introduced. Using the normalized feature vector X_normalized extracted in step S160 as input, a support vector machine (SVM) classifier is trained. This classifier is trained to distinguish the real landing moment from other movement stages. Only when both the heuristic decision tree model and the SVM classifier consider it to be the landing moment, is it finally confirmed that the moment is the landing moment. This double verification mechanism can effectively reduce misjudgment and improve the overall recognition accuracy.

[0205] The mathematical expression of the SVM classifier is as follows: f(X)=sign(Σ(αi*yi*K(Xi,X))+b), where αi is the Lagrange multiplier, yi is the label of the training sample, K(Xi,X) is the kernel function (preferably the radial basis function RBF kernel is used here), and b is the bias term. When f(X)>0, it means that the classifier believes that the moment corresponding to X is the moment of landing. The selection of the RBF kernel is based on its good nonlinear mapping ability and its ability to handle complex decision boundaries in the feature space. The training process of SVM includes selecting appropriate kernel function parameters and regularization parameters, which is usually done through cross-validation to strike a balance between model complexity and generalization ability.

[0206] The output of this multi-level heuristic decision tree model is the preliminary recognition result of the landing moment, including a series of possible landing moment time point sequences {t_1, t_2, ..., t_n} and their corresponding confidence scores {C_1, C_2, ..., C_n}. The confidence score comprehensively considers the judgment results of each level of the decision tree and the output probability of the SVM classifier. Specifically, the confidence score can be defined as the weighted sum of the decision tree judgment results and the SVM output probability. These results will be used as candidate points for subsequent precise positioning, rather than the final measurement results.

[0207] Step S180 , collecting environmental factors, and constructing an environmental compensation model based on the environmental factors and the pixel-level dynamic optical flow field sequence.

[0208] Through systematic literature research and expert consultation, five main environmental factors were identified: temperature, humidity, air pressure, surface temperature of the venue, and light intensity. These factors affect sports performance through different mechanisms: temperature affects muscle contraction efficiency and energy metabolism; humidity affects body temperature regulation; air pressure slightly affects jump height; surface temperature of the venue affects friction; and light intensity affects visual perception and camera image quality.

[0209] In order to accurately capture environmental data, a comprehensive environmental monitoring system is deployed around the long jump venue. The system includes: temperature and humidity sensors (installed on the ground, 1 meter and 2 meters in height), barometers, infrared thermometers, illuminance meters and small weather stations. Preferably, all sensors collect data at a frequency of 10Hz, and each data point is accompanied by a millisecond-accurate timestamp to ensure synchronization with the video frame and optical flow field data.

[0210] In one example, the steps of collecting environmental factors and constructing an environmental compensation model based on the environmental factors and pixel-level dynamic optical flow field sequences include the following steps: standardizing the environmental factors; calculating the Pearson correlation coefficient between the standardized environmental factors and the pixel-level dynamic optical flow field sequence; performing principal component analysis on the standardized environmental factors whose Pearson correlation coefficient is greater than a preset value to obtain the environmental principal components; establishing a multivariate regression model with the environmental principal components as independent variables and the pixel-level dynamic optical flow field sequence as the dependent variable; performing a sensitivity analysis by calculating the partial derivatives of the standardized environmental factors whose Pearson correlation coefficient is greater than a preset value to quantify the degree of influence of the standardized environmental factors whose Pearson correlation coefficient is greater than the preset value; and constructing the environmental compensation model based on the multivariate regression model and the degree of influence.

[0211] In one example, the construction of the environmental compensation model also includes the following steps: evaluating the root mean square error of the multivariate regression model, and if it exceeds a preset threshold, adopting a nonlinear model; using the nonlinear model to refit the relationship between the environmental principal components and the pixel-level dynamic optical flow field sequence; performing a sensitivity analysis by calculating the partial derivatives of the environmental principal components with respect to the pixel-level dynamic optical flow field sequence; constructing an environmental deviation estimation function based on the sensitivity analysis results; introducing a time weighting strategy to construct a time weighting function; applying the time weighting function to the environmental deviation estimation function to obtain a time-varying environmental compensation model; and using the time-varying environmental compensation model to correct the pixel-level dynamic optical flow field sequence.

[0212] Specifically, after the environmental data is collected, all environmental parameters are first standardized:

[0213] Z_j=(X_j-μ_j) / σ_j

[0214] Among them, X_j is the original data of the jth environmental factor (such as temperature, humidity, etc.), μ_j is its average value, and σ_j is its standard deviation. j=1, 2, ..., 5, corresponding to five environmental factors. This standardization process eliminates the influence of different dimensions and makes the influence of each environmental factor directly comparable.

[0215] Next, the Pearson correlation coefficient between the standardized environmental parameter Z_j and the pixel-level dynamic optical flow field sequence F_original(t) is calculated to analyze the relationship between environmental factors and motion:

[0216] r_ZF_j=Σ((Z_j,i-Z_j)(F_i-F)) / √(Σ(Z_j,i-Z_j)²*Σ(F_i-F)²)

[0217] Here, Z_j,i represents the value of the jth standardized environmental parameter at time i, and F_i represents the average optical flow amplitude at time i. Z_j and F are the average values ​​of Z_j and F respectively.

[0218] The results of this correlation analysis are used to preliminarily screen environmental factors that have a significant impact on the optical flow field. Only environmental factors whose absolute values ​​of correlation coefficients exceed the threshold T_corr (such as 0.3) will be included in subsequent PCA and regression analysis.

[0219] Considering the possible multicollinearity problem between environmental factors, principal component analysis (PCA) was performed on the screened standardized environmental parameter data. The steps of PCA include calculating the covariance matrix, solving the eigenvalues ​​and eigenvectors, and selecting the first m principal components that can explain 95% of the variance. These principal components are linear combinations of the original environmental parameters:

[0220] E_i(t)=a_i1*Z_1(t)+a_i2*Z_2(t)+...+a_i5*Z_5(t)

[0221] Where a_ij is the coefficient of the jth environmental factor Z_j in the i-th principal component E_i, that is, the component of the eigenvector vi_i.

[0222] The principal component of the environment E_i(t) obtained by PCA is used as the independent variable, and the pixel-level dynamic optical flow field sequence F_original(t) is used as the dependent variable to fit the multivariate regression model:

[0223]

[0224] Where β_i is the regression coefficient and ε(t) is the error term. Preferably, 10-fold cross validation is used to evaluate the generalization ability of the model. If the root mean square error (RMSE) exceeds the preset threshold T_RMSE (such as 10% of the mean amplitude of the optical flow field), a nonlinear model such as polynomial regression (up to 3rd order) or support vector regression (using RBF kernel) is adopted. In this case, the relationship between the environmental principal component E_i(t) and the pixel-level dynamic optical flow field sequence F_original(t) is refitted using the selected nonlinear model. This refitting process can be expressed as: F_original(t) = f(E_1(t), E_2(t), ..., E_m(t)) + ε(t)

[0225] Where f is a selected nonlinear function, which can be a polynomial function or a function defined by support vector regression.

[0226] By calculating the partial derivatives Conduct sensitivity analysis to quantify the impact of each environmental principal component:

[0227]

[0228] The results of this analysis are used to determine the weights of each environmental principal component in the compensation model.

[0229] Based on the above analysis, an environmental compensation model is constructed:

[0230] F_corrected(t)=F_original(t) / (1+Δ(E(t)))

[0231] Where Δ(E(t)) is the deviation estimation function caused by environmental factor E, and its specific form is:

[0232] Δ(E(t))=w_1*g_1(E_1(t))+w_2*g_2(E_2(t))+...+w_m*g_m(E_m(t))

[0233] Here, g_i is a nonlinear function (can be a polynomial or exponential function) for the i-th principal component, determined by the data fit. w_i is a weight, proportional to the sensitivity S_i: w_i=S_i / Σ_jS_j.

[0234] In order to deal with the time-varying nature of environmental factors, a time-weighted strategy is introduced:

[0235] Δ(E,t)=Σ_i[w_i(t)*g_i(E_i(t))]

[0236] The time weight function w_i(t) uses a Gaussian kernel, with the center set at the moment of take-off and landing, and a standard deviation of 0.1 seconds:

[0237] w_i(t)=exp(-(t-t_takeoff)² / (2*0.1²))+exp(-(t-t_landing)² / (2*0.1²))

[0238] This design reflects the critical impact of environmental factors at the moment of take-off and landing on measurement accuracy.

[0239] Finally, this environmental compensation model improves the accuracy of standing long jump measurement by quantifying the impact of environmental factors on the original dynamic optical flow field. The model parameters directly reflect the degree of influence of different environmental factors on long jump performance, providing teachers with quantitative environmental adaptation guidance. For example, the model may reveal that students perform best within a specific temperature range, or that a certain humidity level has a significant impact on long jump distance.

[0240] Step S190, processing potential landing instant frames and pixel-level dynamic optical flow field sequences based on the environment compensation model, obtaining landing key frames, and calculating the student's long jump score based on the landing key frames.

[0241] In one example, the steps of processing potential landing instant frames and pixel-level dynamic optical flow field sequences based on an environmental compensation model, obtaining landing key frames, and calculating students' long jump scores based on the landing key frames include: correcting the pixel-level dynamic optical flow field sequence based on the environmental compensation model within a preset time window; locating the potential landing instant frames based on the corrected pixel-level dynamic optical flow field sequence using a multi-scale analysis method; searching for the local maximum value of the second comprehensive indicator around the positioning result using a peak detection algorithm, and taking the local maximum value as the landing key frame; converting the two-dimensional coordinates of the landing key frame into three-dimensional spatial coordinates of landing; and calculating the students' long jump scores based on the three-dimensional spatial coordinates of landing and the three-dimensional spatial coordinates of take-off.

[0242] First, use the time points {t_1, t_2, ..., t_n} and their corresponding confidence scores {C_1, C_2, ..., C_n} output in step S170 as the starting point. Preferably, a confidence threshold T_conf (such as 0.7) is set, and t_k is considered as a valid candidate landing moment only when C_k>T_conf (k=1, 2, ..., n). Then, for each valid candidate landing moment t_k, search within a small preset time window [t_k-Δt, t_k+Δt] around it, where Δt is usually set to 5-10 frames.

[0243] In the preset time window, the environment compensation model constructed in step S180 is applied to correct the pixel-level dynamic optical flow field sequence. The specific application process is as follows:

[0244] 1. For each frame within the preset time window, collect the current environmental parameters, including temperature, humidity, air pressure, site surface temperature, and light intensity.

[0245] 2. Input these environmental parameters into the environmental compensation model constructed in step S180 to calculate the corresponding compensation factors.

[0246] 3. Use the calculated compensation factor to correct the pixel-level dynamic optical flow field sequence to obtain F_corrected(t).

[0247] This correction process takes into account the combined effects of multiple environmental factors and improves the accuracy of subsequent analysis by making appropriate adjustments to the original optical flow field. In particular, in outdoor long jump tests, this step can effectively reduce the impact of factors such as wind speed and temperature changes on the measurement results.

[0248] Based on the corrected optical flow field, a multi-scale analysis method is further used to refine the positioning at the moment of landing. In the range of [t_i-Δt, t_i+Δt], the divergence and curl of the optical flow field are calculated at the original resolution, 1 / 2 resolution, and 1 / 4 resolution respectively. Here, the divergence is expressed as , the curl is expressed as , where u and v are the horizontal x and vertical y components of the optical flow field F_corrected(t), respectively. It is worth noting that the divergence reflects the degree of expansion or contraction of the motion, while the curl reflects the intensity of the rotational motion.

[0249] Subsequently, a new comprehensive index J'(t) (second comprehensive index) is defined to quantify the landing characteristics: J'(t)=α*|dv_body(t) / dt|+β*(|dA_body(t) / dt| / A_body(t))+γ*C(t)+δ*|divergence(t)|+ε*|curl(t)|. Wherein, v_body(t) represents the average speed of the student's body, A_body(t) represents the body projection area, C(t) is the confidence score calculated in step S170, and α, β, γ, δ, and ε are weight coefficients.

[0250] Next, a peak detection algorithm is used to search for the local maximum of J'(t) around each candidate landing moment to determine the most likely landing moment t_land. Specifically, t_land=argmax_{t∈[t_i-Δt, t_i+Δt]}J'(t). After determining the landing key frame, the next step is to accurately extract the coordinates of the grounding point of the student's foot. To this end, image processing techniques, such as Canny edge detection and contour analysis, are used to locate the contact point between the foot and the ground. Taking into account the possible deformation of the foot at the moment of landing, the grounding point is preferably defined as the first contact point between the foot contour and the ground. In order to accurately locate this grounding point, sub-pixel edge positioning technology is used in a two-dimensional image coordinate system. By fitting a quadratic curve near the contact area and calculating the intersection of the curve with the horizon, the grounding point coordinates (x_land, y_land) with an accuracy better than a single pixel can be obtained.

[0251] Subsequently, the two-dimensional ground point coordinates need to be converted into three-dimensional space coordinates. This requires back-projection using the camera's intrinsic parameter matrix K and extrinsic parameter matrix [R|t] (R is the rotation matrix and t is the translation vector). The camera's intrinsic parameter matrix K is obtained by calibration before the system is deployed, while the extrinsic matrix is ​​obtained by placing calibration objects in the field and performing PnP solution. The specific back-projection process can be expressed as: X_land=R^(-1)*(λ*K^(-1)*[x_land;y_land;1]-t), where X_land is the three-dimensional space coordinate and λ is the depth factor. In order to determine the depth factor λ, it is preferred to use the plane constraint of the long jump field, substitute this constraint into the above equation, and solve for the value of λ.

[0252] Similarly, a similar method is used to locate the three-dimensional space coordinates X_take of the take-off point. The take-off point is usually pre-marked, and its two-dimensional coordinates can be directly extracted from the video frame, and then its three-dimensional coordinates are obtained using the same back-projection process. Finally, the standing long jump result is obtained by calculating the horizontal distance between the take-off point and the landing point in three-dimensional space: Distance = √((X_land[0]-X_take[0])²+(X_land[1]-X_take[1])²), where X_land[0], X_land[1] and X_take[0], X_take[1] are the horizontal coordinates of the landing point and the take-off point respectively.

[0253] Finally, the system automatically generates a comprehensive report, including long jump results, key moment analysis, technical evaluation, improvement suggestions, environmental factors, historical comparison, etc. This report not only provides students with immediate feedback, but also provides coaches and physical education teachers with detailed technical analysis tools to help develop more targeted training plans and teaching strategies.

[0254] The present invention relates to an automatic recognition method for the landing moment of a standing long jump based on optical flow analysis, which comprises the following steps: using a high-speed camera to collect a high-frame rate video sequence of a student during the entire jumping process; arranging the high-speed camera in the long jump landing area of ​​the student according to preset rules; sequentially performing distortion correction processing, noise suppression processing, contrast optimization processing, white balance processing and gamma correction processing on the high-frame rate video sequence to obtain a high-frame rate video optimization sequence; based on the high-frame rate video optimization sequence, calculating a pixel-level dynamic optical flow field sequence between adjacent frames of the high-frame rate video optimization sequence; recursively decomposing each optical flow field in the pixel-level dynamic optical flow field sequence to obtain a series of different motion The method comprises the following steps: first, collecting motion components of different scales; calculating the complexity index time series corresponding to the motion components; identifying and extracting the key motion features characterizing the student's landing moment based on the complexity index time series; processing the key motion features based on a multi-level heuristic decision tree model to preliminarily identify potential landing moment frames; collecting environmental factors, and constructing an environmental compensation model based on the environmental factors and pixel-level dynamic optical flow field sequences; processing potential landing moment frames and pixel-level dynamic optical flow field sequences based on the environmental compensation model to obtain landing key frames, and calculating the student's long jump results based on the landing key frames, thereby achieving efficient automation throughout the process, comprehensively and meticulously analyzing the movements, and achieving accurate and objective measurement of results.

[0255] This application significantly improves the accuracy and objectivity of standing long jump performance measurement through high-speed camera capture (step S110), dynamic optical flow analysis (step S130) and environmental compensation model (step S180). High frame rate video and precise image processing technology can accurately capture the moment of landing, eliminating subjective factors in manual measurement. The environmental compensation model ensures that the measurement results under different conditions remain consistent and comparable, providing a reliable basis for fair evaluation of student performance.

[0256] The entire process from video acquisition to score calculation is automated, greatly improving test efficiency. Image preprocessing in step S120, motion analysis in step S130-S130, and precise positioning and score calculation at the moment of landing in step S1910 all require no human intervention. This allows teachers to focus on observing and guiding students rather than manually measuring and recording. The system can process data quickly and give scores in real time, speeding up the test process while allowing students to instantly understand their performance.

[0257] Through the dynamic optical flow field calculation in step S130, the recursive decomposition technology in step 4, and the sample entropy calculation in step 5, the system achieves a comprehensive and detailed analysis of the jumping action. This multi-scale and multi-level analysis capability enables teachers to deeply understand students' jumping techniques and identify movement details and problems that are difficult to detect with the naked eye. For example, the system can detect the distribution of force during take-off, the stability of the posture in the air, or the cushioning effect during landing. These detailed analysis results provide teachers with precise diagnostic tools, enabling them to give more targeted technical guidance and improvement suggestions based on the specific situation of each student.

[0258] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0259] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 2 As shown. The computer device includes a processor, a memory, a network interface and a database connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store long jump video data. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for extracting key frames and estimating performance of standing long jump based on sparse representation is implemented.

[0260] Those skilled in the art will understand that Figure 2 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0261] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0262] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0263] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the patent application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent application shall be subject to the attached claims.

Claims

1. A method for automatically identifying the landing moment of standing long jump based on optical flow analysis, characterized in that: The following steps are involved: A high-speed camera is used to collect a high-frame rate video sequence of the student during the entire jumping process; the high-speed camera is arranged in the long jump landing area of ​​the student according to a preset rule; Preprocessing the high frame rate video sequence to obtain a high frame rate video optimized sequence; Based on the high frame rate video optimization sequence, calculating a pixel-level dynamic optical flow field sequence between adjacent frames of the high frame rate video optimization sequence; Recursively decomposing each optical flow field in the pixel-level dynamic optical flow field sequence to obtain a series of motion components at different motion scales; Calculating a complexity index time series corresponding to the motion component; Based on the complexity index time series, identifying and extracting key motion features characterizing the moment when the student lands; Processing the key motion features based on a multi-level heuristic decision tree model to preliminarily identify potential landing moment frames; Collecting environmental factors, and constructing an environmental compensation model based on the environmental factors and the pixel-level dynamic optical flow field sequence; The potential landing moment frame and the pixel-level dynamic optical flow field sequence are processed based on the environmental compensation model to obtain a landing key frame, and the long jump score of the student is calculated based on the landing key frame.

2. The method for automatically identifying the landing moment of standing long jump based on optical flow analysis according to claim 1 is characterized in that: The step of calculating a pixel-level dynamic optical flow field sequence between adjacent frames of the high frame rate video optimization sequence based on the high frame rate video optimization sequence comprises the steps of: For each pair of adjacent frames in the high frame rate video optimization sequence, a 5-layer pyramid is constructed; the resolution of a next layer in the 5-layer pyramid is half of the resolution of a previous layer; The five-layer pyramid is iteratively optimized three times from top to bottom to obtain the pixel-level dynamic optical flow field sequence; The pixel-level dynamic optical flow field sequence is optimized using mirror filling technology and forward-backward consistency check.

3. The method for automatically identifying the landing moment of standing long jump based on optical flow analysis according to claim 1 is characterized in that: The step of recursively decomposing each optical flow field in the pixel-level dynamic optical flow field sequence to obtain a series of motion components of different motion scales includes the following steps: Based on the pixel-level dynamic optical flow field sequence, construct a Gaussian pyramid; The upper layer in the Gaussian pyramid is the result of the next low-pass filtering and downsampling; Calculating the difference between adjacent layers of the Gaussian pyramid to construct a Laplacian pyramid; the Laplacian pyramid represents detail information of a specific scale; The top layer of the Gaussian pyramid and all layers of the Laplacian pyramid are decomposed to obtain the motion components of a series of different motion scales.

4. The method for automatically identifying the landing moment of standing long jump based on optical flow analysis according to claim 1 is characterized in that: The step of calculating the complexity index time series corresponding to the motion component includes the steps of: Calculate the horizontal direction sample entropy time series and the vertical direction sample entropy time series of the motion component; Calculating a time series of complexity ratios between different motion scales corresponding to the motion components; The complexity index time series is acquired based on the horizontal sample entropy time series, the vertical sample entropy time series and the complexity ratio time series.

5. The method for automatically identifying the landing moment of standing long jump based on optical flow analysis according to claim 1 is characterized in that: The step of identifying and extracting key motion features characterizing the moment when the student lands on the ground based on the complexity index time series includes the following steps: Analyzing the complexity index time series to obtain characteristic information corresponding to different motion scales; Based on the characteristic information, design a first comprehensive indicator sequence; Using a peak detection algorithm, and based on time interval constraints, complexity pattern constraints, and continuity constraints, a local maximum value is found in the first comprehensive indicator sequence as the landing instant frame; A key feature vector is extracted from the landing moment frame, and the key feature vector is normalized to obtain the key motion feature characterizing the landing moment of the student.

6. The method for automatically identifying the landing moment of standing long jump based on optical flow analysis according to claim 1 is characterized in that: The step of processing the key motion features based on a multi-level heuristic decision tree model to preliminarily identify potential landing moment frames includes the following steps: Constructing the multi-level heuristic decision tree model, wherein the multi-level heuristic decision tree model is used to process the extracted key motion features; In the first level, comprehensive indicators in the key motion features are evaluated, and moments above a preset threshold are marked as preliminary potential landing moments; In the second level, for the preliminary potential landing moment, the complexity change rate of different scales in the key motion features is analyzed to further screen and obtain the secondary potential landing point; In the third level, for the secondary potential landing points, the changes in the complexity ratios between different scales in the key motion features are evaluated to obtain high-level potential landing points; In the fourth level, for the high-level potential landing points, a dynamic time warping algorithm is used to calculate the similarity between their timing patterns and the predefined typical landing patterns to determine the initial candidate landing moments; global time constraints are applied to the initial candidate landing moments to ensure that there is a sufficient time interval between adjacent initial candidate landing moments to obtain the candidate landing moments; a support vector machine classifier is used to verify the candidate landing moments to further improve recognition accuracy; and the time point sequence of the candidate landing moments that have been screened and verified at all levels and the confidence scores corresponding to the candidate landing moments are output.

7. The method for automatically identifying the landing moment of standing long jump based on optical flow analysis according to claim 1 is characterized in that: The step of collecting environmental factors and constructing an environmental compensation model based on the environmental factors and the pixel-level dynamic optical flow field sequence includes the following steps: Standardizing the environmental factors; Calculating the Pearson correlation coefficient between the standardized environmental factors and the pixel-level dynamic optical flow field sequence; Performing principal component analysis on the standardized environmental factors whose Pearson correlation coefficient is greater than a preset value to obtain environmental principal components; A multivariate regression model is established with the environmental principal component as an independent variable and the pixel-level dynamic optical flow field sequence as a dependent variable; A sensitivity analysis is performed by calculating the partial derivative of the standardized environmental factor whose Pearson correlation coefficient is greater than a preset value, so as to quantify the influence of the standardized environmental factor whose Pearson correlation coefficient is greater than a preset value; Based on the multiple regression model and the impact degree, the environmental compensation model is constructed.

8. The method according to claim 7, characterized in that The construction of the environmental compensation model also includes the following steps: Evaluate the root mean square error of the multivariate regression model and adopt a nonlinear model if it exceeds a preset threshold; Refitting the relationship between the principal components of the environment and the pixel-level dynamic optical flow field sequence using the nonlinear model; Sensitivity analysis is performed by calculating the partial derivatives of the principal components of the environment with respect to the pixel-level dynamic optical flow field sequence; Based on the sensitivity analysis results, an environmental deviation estimation function is constructed; a time weighting strategy is introduced to construct a time weighting function; Applying the time weight function to the environmental deviation estimation function to obtain a time-varying environmental compensation model; The pixel-level dynamic optical flow field sequence is corrected using the time-varying environment compensation model.

9. The method for automatically identifying the landing moment of standing long jump based on optical flow analysis according to claim 1, characterized in that: The step of processing the potential landing moment frame and the pixel-level dynamic optical flow field sequence based on the environmental compensation model, obtaining a landing key frame, and calculating the student's long jump score based on the landing key frame includes: Correcting the pixel-level dynamic optical flow field sequence based on an environmental compensation model within a preset time window; Based on the corrected pixel-level dynamic optical flow field sequence, a multi-scale analysis method is used to locate the potential landing moment frame; A peak detection algorithm is used to search for a local maximum value of the second comprehensive indicator around the positioning result, and the local maximum value is used as the landing key frame; Converting the two-dimensional coordinates of the landing key frame into the three-dimensional space coordinates of the landing; The long jump result of the student is calculated based on the landing three-dimensional space coordinates and the take-off three-dimensional space coordinates.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 9 are implemented.

Images