Athlete training video image editing method based on deep learning

Abstract

The invention relates to the technical field of computer vision and video processing, and discloses a deep learning-based athlete training video image editing method, which comprises the following steps of: analyzing an original code stream of a video to be edited in real time, extracting a compressed domain feature, calculating a motion activity index by utilizing a sliding window, and determining a motion candidate interval; performing sparse decoding on the action candidate interval, extracting a human body key point track by using a convolutional neural network, and constructing a motion track matrix; inputting the motion trail matrix into a time sequence recognition network, generating an action phase sequence under the constraint of a biomechanical phase, and determining a preliminary semantic boundary; the technical scheme of biomechanical phase constraint and time sequence recognition is adopted, the technical effect of accurately locking the logic boundary of the action is achieved, complete coverage of the complete cycle of the action is achieved, and the defects that in a traditional method, due to environment interference or a fixed window, the end of the action is broken, and starting and ending points are unstable are overcome.

Description

Technical Field

[0001] This invention relates to the field of computer vision and video processing technology, specifically to an image editing method for athlete training videos based on deep learning. Background Technology

[0002] In grassroots sports training or school sports settings, coaches and athletes typically use mobile devices such as smartphones to shoot training videos from a single camera. Due to the complex shooting environment and the fact that the videos are mostly in low-bitrate compressed formats, the challenge lies in how to quickly, accurately, and losslessly identify and extract complete video clips of key movements from lengthy raw footage containing a large number of non-critical movements.

[0003] Currently, existing automatic video editing solutions mainly rely on full video decoding combined with motion recognition algorithms to locate keyframes. The typical approach is as follows: fully decode the original video to extract image features or human skeleton information from each frame; use a pre-set motion recognition model to find the peak of motion confidence; extend a fixed window length before and after the peak to determine the editing range; re-encode the determined range and export the edited target short video.

[0004] However, existing technologies still have significant limitations in practical applications: First, boundary positioning is unstable and limited by complex environmental interference. Methods based on fixed windows or simple threshold judgments are prone to causing cut-offs in the action cycle, failing to guarantee the integrity of the start and end points of the action. Second, processing efficiency is low and image quality is degraded. The existing "decoding-recognition-re-encoding" process has a huge computational load on mobile devices, high processing latency, and the clarity of the edited video is reduced due to secondary compression, resulting in the loss of key motion details. To solve the above problems, this invention proposes an image editing method for athlete training videos based on deep learning. By combining temporal boundary detection and video coding domain feature analysis using deep learning, it achieves accurate positioning and low-latency lossless editing of key action boundaries in complex environments, providing an efficient and high-quality motion feedback tool for grassroots sports training. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a deep learning-based image editing method for athlete training videos, thereby resolving the problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an image editing method for athlete training videos based on deep learning, comprising:

[0007] The system analyzes the original bitstream of the video to be edited in real time and extracts the compressed domain features. It then uses a sliding window to calculate the motion activity index and determine the action candidate interval.

[0008] Sparse decoding is performed on the action candidate interval, and the trajectory of human key points is extracted using a convolutional neural network to construct a motion trajectory matrix.

[0009] The motion trajectory matrix is ​​input into the temporal recognition network to generate an action phase sequence under biomechanical phase constraints, thereby determining the initial semantic boundary.

[0010] Based on the initial semantic boundary, bidirectional expansion is performed to determine the candidate segmentation point set and identify the coding frame type and image group structure corresponding to each candidate segmentation point.

[0011] By utilizing the coded frame type and image group structure, a joint optimization objective function including phase integrity loss and recoding cost is constructed, and the comprehensive score of each point in the candidate segmentation point set is calculated.

[0012] Select the optimal value from the overall score, and determine the target clipping boundary based on this optimal value;

[0013] Based on the target clip boundaries, perform stream copying or local recoding to generate the target action segment;

[0014] It captures the boundary fine-tuning signals of the target action segment and updates the weight parameters in the joint optimization objective function through a weak supervision feedback mechanism.

[0015] Preferably, the compressed domain features include motion vectors, residual energy, and macroblock prediction modes;

[0016] The steps In this process, the motion activity index is calculated using a sliding window, which includes: obtaining the intra-prediction ratio of macroblocks in the original bitstream and the magnitude of the motion vector; performing a weighted summation of the intra-prediction ratio and the magnitude of the motion vector within a preset time window; and determining the corresponding bitstream time period as an action candidate interval when the summation result continuously exceeds the activity threshold.

[0017] Preferably, the biomechanical phase constraint includes a phase sequence consistency constraint and a minimum phase duration constraint; the phase sequence consistency constraint is achieved by verifying whether the action phase sequence conforms to a preset kinetic anatomy logical order.

[0018] The specific implementation of the biomechanical phase constraint includes: pre-establishing a standard action timing template that includes a preparation phase, an exertion phase, an airborne phase, and a recovery phase; matching the action phase sequence with the standard action timing template; and correcting the preliminary semantic boundary to include the video frame corresponding to the missing phase when there is a phase loss or phase reversal in the action phase sequence.

[0019] Preferably, the joint optimization objective function The calculation formula is:

[0020] ,

[0021] in, As candidate split points, For phase integrity loss, The cost of recoding and These are weight parameters;

[0022] The phase integrity loss The method for determining the candidate split points is as follows: calculate the candidate split points. The frame difference between the candidate segmentation point and the initial semantic boundary is input into a preset linear penalty function to obtain the coverage deviation value of the candidate segmentation point for the action key feature point.

[0023] The recoding cost The candidate segmentation point is determined based on the displacement distance of the candidate segmentation point from the nearest keyframe in the image group to which it belongs.

[0024] Preferably, the steps In the process, when the target clip boundary hits a keyframe, stream copying is performed directly; when the target clip boundary is located in a non-keyframe, local recoding is performed on the image group containing the target clip boundary.

[0025] The execution steps of the local recoding include: locating the starting position of the image group where the target clip boundary is located, decoding the video frames from the starting position to the target clip boundary, performing secondary compression encoding on the decoded video frames using the original encoding parameters, and splicing them with the bitstream of the subsequent stream copy;

[0026] The target action segment encapsulates structured metadata containing action integrity scores and phase interval labels.

[0027] Preferably, the weakly supervised feedback mechanism includes: when the displacement distance generated by the boundary fine-tuning signal exceeds a preset threshold, increasing the weight parameter corresponding to the phase integrity loss in the joint optimization objective function. The value;

[0028] The weakly supervised feedback mechanism includes: extracting video frame features associated with the boundary fine-tuning signal, using the video frame features as negative samples to input the temporal recognition network for incremental training, and adjusting the connection weights of each neuron in the temporal recognition network.

[0029] Preferably, the steps In this process, constructing the joint optimization objective function includes: establishing a decision matrix composed of several candidate split points, using the Lagrange multiplier method to solve for the global minimum value of the decision matrix under the constraints of the weight parameters, and mapping the coordinates of the split points corresponding to the global minimum value to the comprehensive score.

[0030] Preferably, the sparse decoding includes: determining a sampling step size based on the duration of the action candidate interval, skipping non-reference frames and performing partial decoding according to the sampling step size, to obtain a sparse image frame sequence containing key motion information;

[0031] The extraction of human key point trajectories includes: locating the joint coordinates of the athlete's torso and limbs in the sparse image frame sequence, and using the Kalman filter algorithm to smooth the joint coordinates between adjacent sparse image frames to generate a continuous motion trajectory matrix.

[0032] Preferably, the bidirectional expansion includes: expanding to the left and to the right with a preset expansion step size, centered on the initial semantic boundary. To construct the candidate splitting point set;

[0033] The process of identifying the encoded frame type and image group structure corresponding to each candidate segmentation point includes: parsing the header information of the bitstream, obtaining the frame index value of each candidate segmentation point and the length information of the image group, and determining whether the candidate segmentation point has the physical conditions to perform stream copying.

[0034] Preferably, the adaptive update of the weight parameters includes: calculating the statistical variance of multiple sets of boundary fine-tuning signals, and determining the weight parameters when the statistical variance is lower than a preset convergence threshold. and The globally optimal ratio is determined and then incorporated into the joint optimization objective function as a dedicated editing strategy for this type of sport.

[0035] This invention provides an image editing method for athlete training videos based on deep learning. It has the following beneficial effects:

[0036] 1. This invention adopts a biomechanical phase constraint and timing recognition technology solution to achieve the technical effect of accurately locking the action logic boundary, realizing complete coverage of the entire action cycle, and solving the shortcomings of traditional methods that cause action to be cut off or have unstable start and end points due to environmental interference or fixed windows.

[0037] 2. This invention adopts a compressed domain feature initial screening and sparse decoding technology to achieve a significant reduction in computing power overhead, realize ultra-fast response in mobile environments, and solve the shortcomings of traditional full-data decoding process with high computational load and high processing latency.

[0038] 3. This invention adopts a joint optimization objective function and streaming copy editing technology to achieve the technical effect of coordinating semantic boundaries and coding structure, realize lossless video output, and solve the shortcomings of traditional recoding process that cause image quality loss and loss of motion details.

[0039] 4. This invention adopts a weakly supervised feedback mechanism and a weight adaptive update technology to achieve continuous model evolution and specialized optimization effects, thereby improving editing accuracy through user use and solving the shortcomings of existing solutions in terms of poor versatility and inability to evolve in a closed loop under different motion scenarios. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating the deep learning-based image editing method for athlete training videos according to the present invention. Detailed Implementation

[0041] To enable those skilled in the art to understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention.

[0042] The present invention will now be described in detail with reference to the accompanying drawings:

[0043] Please see the appendix Figure 1 This invention provides an image editing method for athlete training videos based on deep learning, comprising:

[0044] The process involves real-time analysis of the original bitstream of the video to be edited and extraction of compressed domain features. A sliding window is used to calculate motion activity indices and determine candidate motion intervals. Compressed domain features include motion vectors, residual energy, and macroblock prediction modes. (Steps...) In the process of calculating motion activity index using a sliding window, the following steps are taken: obtaining the intra-prediction ratio and motion vector magnitude of macroblocks in the original bitstream; performing a weighted summation of the intra-prediction ratio and motion vector magnitude within a preset time window; and determining the corresponding bitstream time period as an action candidate interval when the summation result continuously exceeds the activity threshold.

[0045] For the action candidate interval, sparse decoding is performed, and the trajectory of human key points is extracted using a convolutional neural network to construct a motion trajectory matrix. The sparse decoding includes: determining the sampling step size according to the duration of the action candidate interval, skipping non-reference frames according to the sampling step size for partial decoding, and obtaining a sparse image frame sequence containing key motion information.

[0046] Extracting the trajectory of key points of the human body includes: locating the coordinates of the joints of the athlete's torso and limbs in a sparse image frame sequence, and using the Kalman filter algorithm to smooth the coordinates of the joints between adjacent sparse image frames to generate a continuous motion trajectory matrix.

[0047] The motion trajectory matrix is ​​input into the temporal recognition network to generate an action phase sequence under biomechanical phase constraints, thus determining the initial semantic boundary. The biomechanical phase constraints include phase sequence consistency constraints and minimum phase duration constraints. The phase sequence consistency constraints are achieved by verifying whether the action phase sequence conforms to the preset motion anatomy logical order.

[0048] The specific implementation of biomechanical phase constraints includes: pre-establishing a standard action timing template that includes preparation phase, force exertion phase, take-off phase and recovery phase; matching the action phase sequence with the standard action timing template; and correcting the preliminary semantic boundary to include the video frame corresponding to the missing phase when there is a phase missing or phase reversal in the action phase sequence.

[0049] Based on the initial semantic boundary, bidirectional expansion is performed to determine the candidate segmentation point set and identify the coding frame type and image group structure corresponding to each candidate segmentation point.

[0050] Bidirectional expansion includes: expanding to the left and to the right, centered on the initial semantic boundary, with a preset expansion step size. To construct a set of candidate splitting points;

[0051] Identifying the encoded frame type and image group structure corresponding to each candidate segmentation point includes: parsing the header information of the bitstream, obtaining the frame index value and image group length information of each candidate segmentation point, and determining whether the candidate segmentation point has the physical conditions to perform stream copying.

[0052] By utilizing the coded frame type and image group structure, a joint optimization objective function incorporating phase integrity loss and recoding cost is constructed to calculate the comprehensive score of each point in the candidate segmentation point set; the joint optimization objective function... The calculation formula is:

[0053] ,

[0054] in, As candidate split points, For phase integrity loss, The cost of recoding and These are weight parameters;

[0055] Phase integrity loss The method for determining the candidate split points is as follows: calculate the candidate split points. The frame difference between the initial semantic boundary and the target segment is input into a preset linear penalty function to obtain the coverage deviation of the candidate segmentation point for the key action feature points; recoding cost. The candidate segmentation point is determined based on its displacement distance from the nearest keyframe in the image group to which it belongs;

[0056] step In this process, the construction of the joint optimization objective function includes: establishing a decision matrix composed of several candidate split points, using the Lagrange multiplier method to solve for the global minimum value of the decision matrix under the constraint of weight parameters, and mapping the coordinates of the split points corresponding to the global minimum value to a comprehensive score;

[0057] Select the optimal value from the overall score, and determine the target clipping boundary based on this optimal value;

[0058] Based on the target clip boundary, perform stream copy extraction or local recoding to generate the target action segment; when the target clip boundary hits a keyframe, perform stream copy extraction directly; when the target clip boundary is located in a non-keyframe, perform local recoding for the image group containing the target clip boundary.

[0059] The execution steps of local recoding include: locating the starting position of the image group where the target clip boundary is located, decoding the video frames from the starting position to the target clip boundary, performing secondary compression encoding on the decoded video frames using the original encoding parameters, and splicing them with the bitstream of the subsequent stream copy; the target action segment encapsulates structured metadata containing action integrity scores and phase interval labels;

[0060] The boundary fine-tuning signal of the target action segment is captured, and the weight parameters in the joint optimization objective function are updated through a weak supervision feedback mechanism.

[0061] When the displacement distance generated by the boundary fine-tuning signal exceeds the preset threshold, the weight parameter corresponding to the phase integrity loss in the joint optimization objective function is increased. The value;

[0062] Extract video frame features associated with boundary fine-tuning signals, use these video frame features as negative samples to input the temporal recognition network for incremental training, and adjust the connection weights of each neuron in the temporal recognition network.

[0063] Adaptive weight parameter updates include: calculating the statistical variance of multiple sets of boundary fine-tuning signals; and determining the weight parameters when the statistical variance is lower than a preset convergence threshold. and The globally optimal ratio is determined and then incorporated into the joint optimization objective function as a dedicated editing strategy for this type of sport.

[0064] The raw video stream refers to the compressed binary data sequence of the video signal after initial encoding processing, before it has been fully decoded and restored. Because this data stream still retains the compression structure and syntax elements from the encoding process, the processing system can efficiently extract various low-level encoding features from it without significantly increasing the decoding computational burden, providing underlying data support for video analysis tasks.

[0065] Compression domain features refer to the coding unit level information directly parsed from the original video bitstream without pixel-level decoding. They can effectively characterize the motion trend, intensity of motion, and spatial texture complexity of objects in the video content, providing important clues for high-level semantic analysis.

[0066] The Motion Activity Index (MOI) is a quantitative parameter calculated using mathematical modeling methods to objectively measure the intensity level of movement contained in video clips. This index is often used as a preliminary screening mechanism to determine whether there are valid movements in the video content that conform to the behavioral characteristics of athletes.

[0067] Sparse decoding is a video processing strategy optimized for computational efficiency. Its core idea is to perform partial decoding on selected image frames at intervals within a video sequence. By significantly reducing the need for full decoding of consecutive frames, this technique can greatly reduce computational redundancy.

[0068] Human keypoint trajectory refers to the spatiotemporal coordinate sequence formed by stably locating and continuously tracking the main joints of the human body in consecutive image frames using computer vision and deep learning algorithms. This trajectory can accurately describe the movement changes of an athlete's limbs in the time dimension, providing basic data for motion analysis.

[0069] A motion trajectory matrix is ​​a structured representation formed by integrating the coordinate data of several key human body points on a time axis according to a specific topological structure. This matrix can simultaneously preserve spatial correlation and temporal evolution information, and can serve as the input data basis for temporal recognition models to support subsequent motion pattern modeling and recognition.

[0070] Temporal recognition networks are deep learning model architectures specifically designed for processing sequential data, effectively modeling the temporal dynamics inherent in motion trajectory matrices. By capturing the temporal dependencies of actions, this network can identify specific motion action categories and different execution stages, providing support for semantic analysis.

[0071] Biomechanical phase constraints are constraints on the sequence and temporal continuity of movements defined based on the principles of human biomechanics. These constraints can eliminate abnormal movement jumps that do not conform to the rationality of human movement during algorithm recognition, significantly improving the physical rationality and reliability of the recognition results.

[0072] Preliminary semantic boundaries refer to the time point coordinates that are initially inferred by the temporal recognition network to identify the start and end positions of an action, reflecting the algorithm's initial estimate of the start and end positions of the action at the physiological level.

[0073] The candidate segmentation point set consists of several time points within a certain temporal neighborhood near the initial semantic boundary, selected according to a preset strategy, that are likely to be the final editing positions. This set forms the basis for refined screening, ensuring that the optimal segmentation position is not missed.

[0074] The joint optimization objective function is a mathematical modeling tool that integrates action segment coverage integrity with the cost of video coding operations. This function searches through multiple sets of candidate segmentation points to find the optimal balance between coverage and computational cost, thereby optimizing overall performance.

[0075] Phase integrity loss is an indicator used to quantify the degree of missing action segments caused by the final cut point deviating from the theoretical action boundary.

[0076] Re-coding cost refers to the additional computational load and potential image quality loss incurred due to partial re-coding caused by clip points not aligning with the keyframe structure in the video encoding. This cost is used to evaluate the overall impact of editing operations on system resources and output quality, and is an important consideration in optimization objectives.

[0077] Stream copying is a video editing method that directly copies and splices data at the compressed bitstream level. Because it completely avoids decoding and re-encoding, this method enables real-time video generation while maintaining the original image quality, making it particularly suitable for applications with high real-time requirements.

[0078] Local recoding is a technique that performs limited decoding and recoding operations on the image group containing the clipping point. This method is suitable for achieving precise clipping at non-keyframe locations while minimizing the impact on overall system performance and image quality, achieving a good balance between accuracy and efficiency.

[0079] Weakly supervised feedback is a learning method that uses users' slight manual adjustments to the results of automatic editing as implicit annotation signals.

[0080] By parsing the raw bitstream in real time and extracting compressed domain features, a large number of static or invalid training background images can be quickly filtered out with extremely low computational overhead without performing full video decoding. Utilizing a sliding window to calculate motion activity metrics improves the sensitivity of capturing motion candidate regions, ensuring accurate identification of the moment an athlete begins to move, even in complex lighting or cluttered backgrounds. This pre-screening mechanism significantly alleviates the memory and computational pressure on mobile devices, reserving ample computational resources for deep processing and improving the overall editing workflow's responsiveness.

[0081] A method combining sparse decoding with convolutional neural networks to extract keypoint trajectories effectively balances motion detail representation with computational efficiency. By constructing a motion trajectory matrix and applying Kalman filtering for smoothing, coordinate jump noise caused by hand shake or environmental occlusion during video recording can be eliminated. The trajectory data generated in this way exhibits extremely high consistency and stability, providing high-quality low-level input for motion phase recognition and ensuring the continuity of motion features in the temporal dimension.

[0082] Introducing biomechanical phase constraints into temporal recognition networks, and rigidly embedding the logical sequence of human movement anatomy into the algorithm's decision-making process, effectively avoids phase loss or motion reversal problems caused by prediction biases in deep learning models. Through matching and correction using standard motion temporal templates, the initial semantic boundaries demonstrate statistical accuracy and meet the actual training requirements of sports professionals. This constraint mechanism enhances the system's robustness in recognizing complex, high-speed movements, ensuring the logical rigor of the motion recognition results at a professional level.

[0083] Based on the initial semantic boundaries, bidirectional expansion and coding structure recognition are performed to achieve the first information alignment between the action semantic layer and the underlying video coding layer. By analyzing the frame type and image group information of candidate segmentation points, the system can know in advance which positions can be losslessly edited and which positions require computational compensation. The fusion of this multi-level feature provides comprehensive data support for integrated decision-making, avoiding computational waste caused by blind editing, and is a key prerequisite for achieving high-performance editing path selection.

[0084] A joint optimization objective function incorporating phase integrity loss and recoding costs is constructed, transforming the originally conflicting objectives of motion integrity and processing efficiency into a unified mathematical optimization problem. By utilizing the Lagrange multiplier method to find the global minimum of the decision matrix, the system ensures that the selected cut points do not truncate motion in pursuit of absolute image quality, nor do they incur a huge recoding load in pursuit of motion integrity. This scientific trade-off mechanism makes the editing strategy highly adaptable, enabling it to output the optimal performance solution under different hardware configurations.

[0085] The optimal value of the comprehensive score is selected to determine the target editing boundary, ensuring that each editing decision is the best practice under current technological constraints. This data-driven boundary determination method eliminates the uncertainty caused by manually setting fixed thresholds, making the action boundary calibration process highly objective and consistent. Through this step, the system can accurately locate the time point that meets the coach's observation needs and triggers the system's most efficient execution path, significantly improving the success rate of automated editing.

[0086] Adaptively performing streaming copy or local recoding based on the target clip boundary attributes significantly optimizes the image quality and processing efficiency of video generation. Streaming copy technology ensures second-level output of original quality when keyframes are hit, while local recoding technology ensures computational processing on extremely small areas when precise editing is necessary, avoiding the loss of image details caused by full recoding in traditional solutions. This flexible execution strategy achieves truly lossless editing, meeting the stringent requirements of high-level training for video detail.

[0087] By introducing a weakly supervised feedback mechanism to capture fine-tuning signals, the system acquires learning and evolutionary capabilities similar to those of a human agent. Through statistical variance analysis and automatic parameter fixing, the editing strategy can be personalized and evolved according to the characteristics of different sports and the user's editing habits, significantly reducing the cost of human intervention in long-term system use. The incremental training process continuously adjusts the weight parameters of neurons, making the algorithm increasingly accurate when processing the same type of sports videos, ultimately achieving a spiral improvement in editing accuracy.

[0088] Example 1: Lossless and high-speed editing on mobile devices in an indoor gymnastics training scenario

[0089] In this embodiment, the present invention is specifically applied in an indoor gymnastics training environment, demonstrating efficient and accurate video processing capabilities. The coach uses the camera function of a smartphone to record the athlete's horizontal bar movements from a fixed position. Because gymnastics movements are highly structured, including clearly defined preparatory, power exertion, and take-off phases, the system can perform intelligent recognition and processing based on these phases. The specific steps are as follows:

[0090] The system analyzes the raw video stream in real time and extracts motion vector data from macroblocks in the horizontal bar region. It dynamically monitors changes in the athlete's movements using sliding window technology. When a sharp increase in the motion vector magnitude is detected at the moment of takeoff, and this increase is continuous... When multiple frames of data exceed the preset activity threshold, the system will automatically lock the candidate time interval containing the complete somersault action.

[0091] For the identified candidate regions, the system performs sparse decoding and utilizes a lightweight convolutional neural network to quickly and accurately locate key points on the athlete's body. The generated trajectory matrix is ​​then input into a temporal recognition network. Guided by biomechanical constraints, this effectively eliminates interference from background movement, thus accurately defining the initial semantic boundaries of the tumbling motion.

[0092] The system revolves around the identified semantic boundaries, in the vicinity Searching and performing operations in multiple frames of images Comparison and calculation revealed that selecting the clipping point at a specific frame before the semantic boundary ensures that the loss of action phase integrity approaches zero, while the additional cost of recoding is zero. Based on this, the system ultimately selected this... The frame serves as the boundary of the target clip.

[0093] Once a boundary hits a keyframe, the system directly performs a stream copy operation from the original bitstream to extract the segment; this process is efficient and fast, and the system can... Generate and maintain in a very short time The original quality action footage allows coaches to immediately observe the details of athletes gripping the bar on their phones with lossless clarity, greatly improving the immediacy and accuracy of training feedback.

[0094] Example 2: Precise phase compensation editing of the starting phase in outdoor track and field sprints

[0095] In this embodiment, the present invention is applied to a sprint starting training scenario in an outdoor track and field stadium. Due to the complex environment and the influence of changes in lighting and background crowds, the semantic recognition process is prone to minor deviations. Therefore, the system introduces a local recoding mechanism to ensure the integrity of the actions and the accuracy of the timing sequence. The specific implementation is as follows:

[0096] By analyzing the residual energy of the video, the system effectively eliminates dynamic background interference such as wind blowing through trees, and accurately locates the key time interval from the athlete's stationary state to the start of the race. It adopts a sampling step-size frame-skipping decoding strategy to efficiently construct the motion trajectory matrix of the starting posture, providing data support for subsequent processing.

[0097] The system analyzes the standard sprint motion template and finds that the key action of the push-off and swing is missing in the initial recognition results. Based on this, the semantic boundary is automatically corrected. With the corrected boundary as the center, the system expands the time range in both directions and generates a series of candidate segmentation points. At the same time, it detects that there is a certain distance between the nearest keyframe and the actual starting point of the action.

[0098] After careful consideration, the system found that forcibly using the stream copy method would result in missing starting actions. The time difference between the start and end of the frame results in a significant loss of phase integrity. Therefore, the system selects non-critical frames with relatively low recoding costs and the ability to preserve motion integrity as the segmentation points; after locating the starting position of the image group, local recoding is performed at the start moment, and seamlessly spliced ​​with the stream copy bitstream.

[0099] The system encapsulates various structured metadata, including the start-up response time, in the generated target video clips. This method effectively solves the problem of traditional editing methods easily resulting in cut-off or trimmed parts at the start-up boundary, while ensuring extremely low processing latency. It fully preserves the moment of the start-up burst, contributing to more comprehensive and reliable training and analysis.

[0100] Example 3: Adaptive Evolution of Feedback Algorithms in Complex Ball Game Training

[0101] In this embodiment, the present invention is applied to a basketball shooting training scenario. Given that the endpoint of the shooting motion varies due to different athletes' habits, the system introduces a user feedback mechanism to achieve self-optimization and evolution of the algorithm. The specific steps are as follows:

[0102] The system automatically identifies and cuts out shooting motion segments, uses a temporal recognition network to identify the hand phase under the preset shooting motion logic, and generates a short video containing the entire process from the ball leaving the hand to entering the net as the preliminary processing result.

[0103] After watching the generated video, if the athlete believes the end time is too early, they can manually drag the boundary backward to increase it. Duration in seconds. The system captures the boundary fine-tuning signal in real time and calculates whether the displacement distance exceeds the preset threshold.

[0104] Once the fine-tuning operation triggers the weakly supervised feedback mechanism, the system automatically adjusts the weight of the phase integrity loss in the objective function. Simultaneously, the fine-tuned video frame features are fed back as negative samples to the temporal recognition network for incremental training to optimize the model's recognition capabilities.

[0105] After multiple iterations of shooting video editing and feedback, the system calculates the variance of the fine-tuning operation. When it finds that convergence has been achieved, the optimized parameters are applied. and The configuration ratio is fixed to the athlete's exclusive shooting editing strategy. Afterward, the algorithm can automatically include the complete ball entry and landing process when processing such videos without human intervention.

[0106] Example 1 verifies the processing capability of the present invention under standardized motion logic. This is demonstrated through application in gymnastics horizontal bar training. to When faced with actions exhibiting clear biomechanical phases, the process can quickly locate and accurately hit keyframes using compressed domain features. The technical value of this embodiment lies in: achieving... Instant response within seconds and Zero lossless preservation of original image quality; verification of the superiority of the stream copy capture path in handling high-performance hardware requirements, solving the drawbacks of traditional full-length decoding solutions on mobile devices, such as slow processing speed and high power consumption.

[0107] Example 2 verifies the reliability of the invention under complex environmental interference. Through application in a track and field starting training environment with uneven lighting and cluttered backgrounds, it demonstrates the ability of biomechanical phase constraints to correct semantic boundaries and the support of local recoding for non-key frame segmentation points. The technical value of this example lies in ensuring processing latency is lower than [previous value]. Under the premise of seconds, the phenomenon of cut-off and uncut animation is eliminated; by jointly optimizing the objective function, the conflict between semantic integrity and encoding physical constraints is resolved, and the high-precision editing capability of the system under non-ideal shooting conditions is demonstrated.

[0108] Example 3 verifies the evolutionary capability of the invention under different user habits. This is demonstrated through the capture of fine-tuning signals during basketball shooting training. How can a weakly supervised feedback mechanism guide the algorithm to perform incremental learning? The technical value of this embodiment lies in: realizing the automatic migration of editing strategies from general to specialized and personalized, enabling the model accuracy to increase spirally with usage time; and overcoming the limitations of traditional editing algorithms with fixed logic that cannot adapt to the different movement habits of athletes or the personalized needs of specific events, thus achieving intelligent editing that is personalized for each individual.

[0109] In summary, the deep learning-based athlete training video image editing method proposed in this invention, through the three representative embodiments described above, comprehensively covers the entire technical loop from bottom-level bitstream parsing and mid-level semantic recognition to top-level feedback optimization.

[0110] Embodiments of the present invention have been presented and described. It will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to the embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for image editing of athlete training videos based on deep learning, characterized in that, include: The system analyzes the original bitstream of the video to be edited in real time and extracts the compressed domain features. It then uses a sliding window to calculate the motion activity index and determine the action candidate interval. Sparse decoding is performed on the action candidate interval, and the trajectory of human key points is extracted using a convolutional neural network to construct a motion trajectory matrix. The motion trajectory matrix is ​​input into a temporal recognition network to generate an action phase sequence under biomechanical phase constraints, thereby determining the initial semantic boundary. Based on the initial semantic boundary, bidirectional expansion is performed to determine the candidate segmentation point set and identify the coding frame type and image group structure corresponding to each candidate segmentation point. By utilizing the coded frame type and image group structure, a joint optimization objective function including phase integrity loss and recoding cost is constructed, and the comprehensive score of each point in the candidate segmentation point set is calculated. Select the optimal value from the overall score, and determine the target clipping boundary based on this optimal value; Based on the target clip boundaries, perform stream copying or local recoding to generate the target action segment; It captures the boundary fine-tuning signals of the target action segment and updates the weight parameters in the joint optimization objective function through a weak supervision feedback mechanism.

2. The image editing method for athlete training videos based on deep learning according to claim 1, characterized in that, The compressed domain features include motion vectors, residual energy, and macroblock prediction modes; The steps In this process, the motion activity index is calculated using a sliding window, which includes: obtaining the intra-prediction ratio of macroblocks in the original bitstream and the magnitude of the motion vector; performing a weighted summation of the intra-prediction ratio and the magnitude of the motion vector within a preset time window; and determining the corresponding bitstream time period as an action candidate interval when the summation result continuously exceeds the activity threshold.

3. The image editing method for athlete training videos based on deep learning according to claim 1, characterized in that, The biomechanical phase constraints include phase sequence consistency constraints and shortest phase duration constraints; the phase sequence consistency constraints are implemented by verifying whether the action phase sequence conforms to a preset kinetic anatomy logical order. The specific implementation of the biomechanical phase constraint includes: pre-establishing a standard action timing template that includes a preparation phase, an exertion phase, an airborne phase, and a recovery phase; matching the action phase sequence with the standard action timing template; and correcting the preliminary semantic boundary to include the video frame corresponding to the missing phase when there is a phase loss or phase reversal in the action phase sequence.

4. The image editing method for athlete training videos based on deep learning according to claim 1, characterized in that, The joint optimization objective function The calculation formula is: ， in, As candidate split points, For phase integrity loss, The cost of recoding and These are weight parameters; The phase integrity loss The method for determining the candidate split points is as follows: calculate the candidate split points. The frame difference between the candidate segmentation point and the initial semantic boundary is input into a preset linear penalty function to obtain the coverage deviation value of the candidate segmentation point for the action key feature point. The recoding cost The candidate segmentation point is determined based on the displacement distance of the candidate segmentation point from the nearest keyframe in the image group to which it belongs.

5. The image editing method for athlete training videos based on deep learning according to claim 1, characterized in that, The steps In the process, when the target clip boundary hits a keyframe, stream copying and interception are performed directly; When the target clip boundary is located in a non-keyframe, local recoding is performed on the image group containing the target clip boundary; The execution steps of the local recoding include: locating the starting position of the image group where the target clip boundary is located, decoding the video frames from the starting position to the target clip boundary, performing secondary compression encoding on the decoded video frames using the original encoding parameters, and splicing them with the bitstream of the subsequent stream copy; The target action segment encapsulates structured metadata containing action integrity scores and phase interval labels.

6. The image editing method for athlete training videos based on deep learning according to claim 1, characterized in that, The weakly supervised feedback mechanism includes: when the displacement distance generated by the boundary fine-tuning signal exceeds a preset threshold, increasing the weight parameter corresponding to the phase integrity loss in the joint optimization objective function. The value; The weakly supervised feedback mechanism includes: extracting video frame features associated with the boundary fine-tuning signal, using the video frame features as negative samples to input the temporal recognition network for incremental training, and adjusting the connection weights of each neuron in the temporal recognition network.

7. The image editing method for athlete training videos based on deep learning according to claim 4, characterized in that, The steps In this process, constructing the joint optimization objective function includes: establishing a decision matrix composed of several candidate split points, using the Lagrange multiplier method to solve for the global minimum value of the decision matrix under the constraints of the weight parameters, and mapping the coordinates of the split points corresponding to the global minimum value to the comprehensive score.

8. The image editing method for athlete training videos based on deep learning according to claim 1, characterized in that, The sparse decoding includes: determining the sampling step size based on the duration of the action candidate interval, skipping non-reference frames and performing partial decoding according to the sampling step size, and obtaining a sparse image frame sequence containing key motion information; The extraction of human key point trajectories includes: locating the joint coordinates of the athlete's torso and limbs in the sparse image frame sequence, and using the Kalman filter algorithm to smooth the joint coordinates between adjacent sparse image frames to generate a continuous motion trajectory matrix.

9. The image editing method for athlete training videos based on deep learning according to claim 1, characterized in that, The bidirectional expansion includes: expanding to the left and to the right with a preset expansion step size, centered on the initial semantic boundary. To construct the candidate splitting point set; The process of identifying the encoded frame type and image group structure corresponding to each candidate segmentation point includes: parsing the header information of the bitstream, obtaining the frame index value of each candidate segmentation point and the length information of the image group, and determining whether the candidate segmentation point has the physical conditions to perform stream copying.

10. The image editing method for athlete training videos based on deep learning according to claim 6, characterized in that, The adaptive update of the weight parameters includes: calculating the statistical variance of multiple sets of boundary fine-tuning signals; and determining the weight parameters when the statistical variance is lower than a preset convergence threshold. and The globally optimal ratio is determined and then incorporated into the joint optimization objective function as a dedicated editing strategy for this type of sport.

Images