Low-power-consumption real-time basketball action recognition method and system based on artificial intelligence

Abstract

The invention provides a low-power-consumption real-time basketball action recognition method and system based on artificial intelligence, and relates to the technical field of artificial intelligence, and the method comprises the steps: collecting motion and visual data through a multi-modal sensor, carrying out the time alignment of the data, and decomposing the shooting action into three stages of preparation, force generation and following based on a kinematics constraint model. And respectively extracting feature vectors, executing completion degree detection and triggering evaluation, carrying out attitude scoring calculation on each stage, and finally carrying out weighted fusion to obtain a comprehensive scoring result. According to the method, accurate decomposition and real-time evaluation of the basketball shooting action are realized, the consumption of computing resources is reduced, and the evaluation efficiency and accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a low-power real-time basketball motion recognition method and system based on artificial intelligence. Background Technology

[0002] Basketball, as a global sport, relies heavily on shooting technique as one of its core skills. With technological advancements, analyzing and evaluating basketball shooting motions using modern sensing technology and artificial intelligence algorithms has become an important method for improving athletes' training effectiveness. In both competitive sports and amateur basketball training, accurate assessment and guidance of shooting posture are crucial for improving accuracy and reducing sports injuries.

[0003] Traditional basketball shooting motion analysis relies primarily on coaches' experience and video replays, methods that suffer from high subjectivity and poor real-time performance. In recent years, with the development of sensor technology, computer vision, and machine learning, data-driven motion recognition and evaluation methods have gradually emerged. Existing technologies mainly collect motion data through wearable sensors or visual capture systems, combining this data with artificial intelligence algorithms for motion recognition and quality assessment.

[0004] Existing technologies generally suffer from high computational resource requirements. Most systems need to run complex deep learning models on high-performance computing platforms, making real-time processing on low-power devices difficult and limiting their widespread use in daily training and general consumer applications. Current action recognition methods lack fine-grained breakdown of the basketball shooting motion, typically treating the shooting motion as a whole for recognition and evaluation, ignoring the technical characteristics and importance of the preparation, power generation, and follow-through phases. This results in unspecific evaluation results, failing to provide effective guidance for refining shooting techniques. Existing systems often rely on single-modality data acquisition, using either only visual data or only inertial sensor data, lacking multimodal data fusion mechanisms. This prevents them from fully utilizing the complementary advantages of different sensing methods and makes it difficult to maintain stable recognition accuracy and evaluation reliability in various complex environments. Summary of the Invention

[0005] This invention provides a low-power real-time basketball motion recognition method and system based on artificial intelligence, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a low-power real-time basketball motion recognition method based on artificial intelligence, comprising: Motion and visual data during the basketball shooting process are collected using a multimodal sensor. The motion data includes acceleration and angular velocity signals, and the visual data includes joint position sequences. The motion data and the visual data are time-aligned to obtain a time-synchronized multimodal data sequence; The multimodal data sequence is decomposed based on the kinematic constraint model, and the continuous shooting action is decomposed into the preparation phase, the power generation phase and the follow-up phase to obtain segmented data for each phase. Action feature vectors are extracted from the segmented data of each stage; Perform stage completion detection on the segmented data of each stage. When the segmented data of a certain stage is detected to meet the completion judgment condition, generate an evaluation trigger signal for that stage. The completion judgment condition includes the duration of the stage reaching a preset lower limit and the change in the key features of the stage reaching a preset change range. In response to the evaluation trigger signal, the posture score is calculated on the action feature vector of the corresponding stage to obtain the posture score of that stage. After completing the posture score calculation for all stages, the posture scores of each stage are weighted and fused to obtain the comprehensive posture score. The comprehensive posture score is output as the shooting posture evaluation result.

[0007] The multimodal data sequence is decomposed based on a kinematic constraint model, breaking down the continuous shooting motion into a preparation phase, a power generation phase, and a follow-up phase. The resulting segmented data for each phase includes: Peak detection is performed on the acceleration signal in the multimodal data sequence to identify local maxima and local minima of the acceleration amplitude as motion intensity transition points. Simultaneously, trajectory analysis is performed on the joint position sequence in the multimodal data sequence to identify the moment when the joint motion direction changes in the opposite direction as attitude transition points, thus obtaining a set of candidate boundary points containing motion intensity transition points and attitude transition points. Project the motion intensity conversion points and attitude conversion points in the candidate boundary point set onto a unified time axis, calculate the matching degree between each candidate boundary point and the stage boundary features in the preset kinematic template, and select the candidate boundary points with matching degrees exceeding the preset matching threshold as stage boundary points. The stage boundary points include the first boundary point between the preparation stage and the force exertion stage and the second boundary point between the force exertion stage and the follow stage. Using the first boundary point and the second boundary point as the dividing points, the multimodal data sequence is divided into a first segment of data corresponding to the preparation stage, a second segment of data corresponding to the exertion stage, and a third segment of data corresponding to the following stage, thus obtaining the segmented data for each stage.

[0008] Trajectory analysis is performed on the joint position sequence in the multimodal data sequence to identify the moment when the joint movement direction changes in the opposite direction as the attitude transition point, including: From the joint position sequence, key joints with a correlation degree exceeding a preset correlation threshold with the shooting action are selected. The key joints include the wrist joint, elbow joint, shoulder joint, and knee joint to obtain a key joint position subsequence. The position coordinates of each key joint in the key joint position subsequence at consecutive time points are differentially calculated to obtain the velocity vector sequence of each key joint at each time point. The velocity vectors of each key joint in the velocity vector sequence are subjected to the inner product operation of adjacent moments. When the inner product value of the velocity vector of a certain key joint at a certain moment and the velocity vector at the previous moment changes from a positive value to a negative value, the moment is determined to be the moment of direction reversal of the key joint, and the set of direction reversal moments of each key joint is obtained. The reversal times of different key joints in the set of reversal times are clustered in time. Multiple reversal times with a time difference less than a preset time window are grouped into the same cluster, and the center time of each cluster is selected as the attitude transition point.

[0009] Perform stage completion detection on the segmented data of each stage. When the segmented data of a certain stage is detected to meet the completion judgment condition, generate an evaluation trigger signal for that stage, including: Real-time duration statistics are performed on the segmented data of each stage. By recording the start time and current time of the segmented data of each stage, the cumulative duration of the segmented data of each stage is calculated. The cumulative duration is compared with the preset minimum duration threshold corresponding to the stage. When the cumulative duration reaches the preset minimum duration threshold, a duration marker for the stage is generated. The characteristic fluctuation degree of segmented data in each stage within a continuous time window is monitored, and the variance index of the characteristic fluctuation degree is calculated. When the variance index is lower than the preset stability threshold and remains below the preset stability duration, the stage is determined to have entered a stable state and a stability achievement mark for the stage is generated. When a certain stage simultaneously possesses both the duration target identifier and the stability target identifier, it is determined that the segmented data of that stage meets the completion determination condition, and an evaluation trigger signal containing the stage identifier and completion time information is generated.

[0010] In response to the evaluation trigger signal, a posture score is calculated on the action feature vector of the corresponding stage to obtain the posture score for that stage, including: In response to the evaluation trigger signal, when the stage corresponding to the evaluation trigger signal is the force exertion stage or the follow-up stage, the posture score of the preceding stage of the shooting action and the action feature vector of the preceding stage are extracted from the buffer. When the stage corresponding to the evaluation trigger signal is the preparation stage, the stage is marked as the starting stage and the preceding data extraction is skipped. For non-initial stages, a transition smoothness index is calculated between the action feature vector of the corresponding stage and the action feature vector of the preceding stage. The transition smoothness index is obtained by calculating the continuity measure of the two feature vectors in a preset transition feature space. A coherence correction coefficient is generated based on the transition smoothness index. The motion feature vector of the corresponding stage is compared with the standard posture template of the corresponding stage, the independent posture deviation value of the stage is calculated, and the independent score value of the stage is generated based on the independent posture deviation value. For non-initial stages, the independent score value is multiplied by the coherence correction coefficient and then superimposed with the inherited correction amount calculated based on the attitude score of the preceding stage to obtain the attitude score for that stage. For the initial stage, the independent score value is directly used as the attitude score for that stage.

[0011] Calculate the transition smoothness index between the action feature vector of the corresponding stage and the action feature vector of the preceding stage. The transition smoothness index is obtained by calculating the continuity measure value of the two feature vectors in a preset transition feature space. Generate a coherence correction coefficient based on the transition smoothness index, including: The action feature vectors of the corresponding stage and the action feature vectors of the preceding stage are projected onto the preset transition feature space to obtain the projected feature vectors of the corresponding stage and the projected feature vectors of the preceding stage. Extract the end feature component of the projection feature vector of the corresponding stage in the preset transition feature space and the start feature component of the projection feature vector of the preceding stage in the preset transition feature space, calculate the feature distance between the end feature component and the start feature component, and construct the stage connection deviation based on the feature distance; Perform time-series derivative operations on the projected feature vectors of the corresponding stage and the projected feature vectors of the preceding stage to obtain the feature rate of change vectors of the corresponding stage and the preceding stage. Calculate the directional consistency metric between the feature rate of change vectors of the corresponding stage and the preceding stage, and construct a trend continuity index based on the directional consistency metric. The transition smoothness index is obtained by fusing the stage connection deviation and the trend continuity index; the transition smoothness index is then mapped to a continuity correction coefficient with a value range within a preset interval through coherence mapping.

[0012] A second aspect of the present invention provides a low-power real-time basketball motion recognition system based on artificial intelligence, comprising: The first unit is used to collect motion data and visual data during the basketball shooting process through a multimodal sensor. The motion data includes acceleration signals and angular velocity signals, and the visual data includes joint position sequences. The second unit is used to perform time alignment processing on the motion data and the visual data to obtain a time-synchronized multimodal data sequence; The third unit is used to decompose the multimodal data sequence based on the kinematic constraint model, decomposing the continuous shooting action into a preparation phase, a power generation phase, and a follow-up phase, and obtaining segmented data for each phase. The fourth unit is used to extract action feature vectors from the segmented data of each stage. The fifth unit is used to perform stage completion detection on the segmented data of each stage. When it is detected that the segmented data of a certain stage meets the completion judgment condition, an evaluation trigger signal is generated for that stage. The completion judgment condition includes the duration of the stage reaching a preset lower limit and the change in the key feature of the stage reaching a preset change range. The sixth unit is used to calculate the posture score of the action feature vector of the corresponding stage in response to the evaluation trigger signal, and obtain the posture score of the stage. After completing the posture score calculation of all stages, the posture scores of each stage are weighted and fused to obtain the comprehensive posture score. The seventh unit is used to output the comprehensive posture score as the shooting posture evaluation result.

[0013] A third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0014] Fourth aspect of the present invention, A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] The beneficial effects of this application are as follows: This invention collects motion and visual data during the basketball shooting process using multimodal sensors, and performs time alignment and kinematic constraint-based decomposition processing on the data, thereby achieving precise segmentation and feature extraction of the basketball shooting motion and improving the accuracy and real-time performance of motion recognition.

[0016] This invention introduces a stage completion detection mechanism. When a certain action stage meets the preset completion judgment condition, the evaluation is triggered immediately, avoiding the delay of waiting for the entire action to be completed before evaluation, reducing the system's computing resource consumption, and realizing low-power real-time processing.

[0017] This invention calculates and weights the feature vectors of each stage of the shooting motion to obtain a comprehensive score, providing fine-grained shooting posture evaluation results. This can help users improve the deficiencies in their shooting motion in a targeted manner and effectively improve training results. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the low-power real-time basketball motion recognition method based on artificial intelligence, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the joint position and attitude transition point detection process. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0021] Figure 1 This is a flowchart illustrating the low-power real-time basketball motion recognition method based on artificial intelligence according to an embodiment of the present invention. Figure 1 As shown, the method includes: Motion and visual data during the basketball shooting process are collected using a multimodal sensor. The motion data includes acceleration and angular velocity signals, and the visual data includes joint position sequences. The motion data and the visual data are time-aligned to obtain a time-synchronized multimodal data sequence; The multimodal data sequence is decomposed based on the kinematic constraint model, and the continuous shooting action is decomposed into the preparation phase, the power generation phase and the follow-up phase to obtain segmented data for each phase. Action feature vectors are extracted from the segmented data of each stage; Perform stage completion detection on the segmented data of each stage. When the segmented data of a certain stage is detected to meet the completion judgment condition, generate an evaluation trigger signal for that stage. The completion judgment condition includes the duration of the stage reaching a preset lower limit and the change in the key features of the stage reaching a preset change range. In response to the evaluation trigger signal, the posture score is calculated on the action feature vector of the corresponding stage to obtain the posture score of that stage. After completing the posture score calculation for all stages, the posture scores of each stage are weighted and fused to obtain the comprehensive posture score. The comprehensive posture score is output as the shooting posture evaluation result.

[0022] In one optional implementation, the multimodal data sequence is decomposed based on a kinematic constraint model, breaking down the continuous shooting motion into a preparation phase, a power generation phase, and a follow-up phase, resulting in segmented data for each phase, including: Peak detection is performed on the acceleration signal in the multimodal data sequence to identify local maxima and local minima of the acceleration amplitude as motion intensity transition points. Simultaneously, trajectory analysis is performed on the joint position sequence in the multimodal data sequence to identify the moment when the joint motion direction changes in the opposite direction as attitude transition points, thus obtaining a set of candidate boundary points containing motion intensity transition points and attitude transition points. Project the motion intensity conversion points and attitude conversion points in the candidate boundary point set onto a unified time axis, calculate the matching degree between each candidate boundary point and the stage boundary features in the preset kinematic template, and select the candidate boundary points with matching degrees exceeding the preset matching threshold as stage boundary points. The stage boundary points include the first boundary point between the preparation stage and the force exertion stage and the second boundary point between the force exertion stage and the follow stage. Using the first boundary point and the second boundary point as the dividing points, the multimodal data sequence is divided into a first segment of data corresponding to the preparation stage, a second segment of data corresponding to the exertion stage, and a third segment of data corresponding to the following stage, thus obtaining the segmented data for each stage.

[0023] Multimodal data sequences were acquired by installing inertial sensors and optical markers on key parts of the athlete's body. The data included acceleration signals, angular velocity signals, and three-dimensional joint coordinates. During the acquisition, the athlete performed five standard shooting motions, each lasting approximately three seconds, with a sampling frequency of 200 Hz, resulting in approximately 3000 data points. The acquired data was preprocessed and stored in a multimodal data matrix with dimensions of 3000×25. The matrix contained 25 features, including the three-dimensional coordinates of nine joints and acceleration and angular velocity data acquired by two inertial sensors.

[0024] When performing peak detection on acceleration signals in multimodal data sequences, an adaptive threshold sliding window algorithm is used to identify local extrema. Specifically, the window width is set to 50 sampling points, and the window sliding step size is 10 sampling points. Within each window, the mean μ and standard deviation σ of the wrist acceleration signal are calculated. A threshold is set to μ + 2σ. Points within the window that exceed the threshold and are greater than their adjacent values ​​are identified as local maxima, and points that are below the threshold and are less than their adjacent values ​​are identified as local minima. For example, within the interval of 500-550 sampling points, the wrist acceleration is detected to reach a local maximum of 15.6 m / s² at point 520. 2 It reaches a local minimum of -12.8 m / s at point 535. 2 These two points are marked as the motion intensity transition points.

[0025] Simultaneously, trajectory analysis was performed on the joint position sequence, calculating the displacement vectors of three key joints—wrist, elbow, and shoulder—between consecutive frames. When the direction angle of the displacement vector changed by more than 90 degrees, it was determined that the joint movement direction had reversed, and the corresponding time point was marked as a posture transition point. For example, the displacement vector of the wrist joint at sampling point 515 was (0.12m, 0.28m, 0.05m), while at sampling point 516 it became (-0.08m, 0.30m, 0.04m). The calculated vector angle was 102 degrees, exceeding the 90-degree threshold; therefore, point 516 was marked as a posture transition point. Using the above method, a total of 28 motion intensity transition points and 35 posture transition points were identified in the entire data sequence, forming a candidate boundary point set.

[0026] All points in the candidate boundary point set are projected onto a unified time axis to form a one-dimensional time point sequence. The preset kinematic template contains the phase boundary features of the standard shooting motion, such as the boundary features between the preparation phase and the power phase: wrist acceleration changes from negative to positive, elbow extension begins to accelerate, and shoulder elevation slows down. For each candidate boundary point, the matching degree between the motion parameters within a 50ms window before and after it and the template features is calculated. The matching degree is calculated using a feature-weighted summation method, with wrist acceleration weighted at 0.5, elbow motion weighted at 0.3, and shoulder motion weighted at 0.2. A score of 1 is awarded when the parameter change trend within the window before and after the candidate point is consistent with the template, and a score of 0 is awarded when they are inconsistent. The matching degree is obtained by multiplying each feature score by its corresponding weight and then summing the results.

[0027] When the matching degree exceeds a preset matching threshold of 0.7, the candidate boundary point is determined as a stage boundary point. For example, the matching degree of the 516th sampling point is calculated to be 0.85, which exceeds the threshold of 0.7, and it is determined as the first boundary point between the preparation stage and the power exertion stage; the matching degree of the 748th sampling point is 0.78, and it is determined as the second boundary point between the power exertion stage and the follow-up stage. In this way, two key boundary points are identified in each shooting motion.

[0028] Using the first and second boundary points as the dividing points, the multimodal data sequence is divided into three segments. In the example above, sampling points 1-516 constitute the first segment, corresponding to the preparation stage; sampling points 517-748 constitute the second segment, corresponding to the exertion stage; and sampling points 749-1000 constitute the third segment, corresponding to the following stage. After the segmentation, each segment is stored as an independent data structure, containing the time index, multimodal feature values, and stage label for that stage.

[0029] The decomposition results were manually verified and compared with the manual breakdown results of professional coaches. The time point deviations were all within 100ms, meeting the accuracy requirements for subsequent technical movement evaluation. Experiments show that this method exhibits good stability and accuracy in the decomposition of shooting movements of different athletes, and can effectively support subsequent movement quality evaluation and technical guidance.

[0030] In one optional implementation, trajectory analysis is performed on the joint position sequence in the multimodal data sequence to identify the moment when the joint movement direction reverses as the attitude transition point, including: From the joint position sequence, key joints with a correlation degree exceeding a preset correlation threshold with the shooting action are selected. The key joints include the wrist joint, elbow joint, shoulder joint, and knee joint to obtain a key joint position subsequence. The position coordinates of each key joint in the key joint position subsequence at consecutive time points are differentially calculated to obtain the velocity vector sequence of each key joint at each time point. The velocity vectors of each key joint in the velocity vector sequence are subjected to the inner product operation of adjacent moments. When the inner product value of the velocity vector of a certain key joint at a certain moment and the velocity vector at the previous moment changes from a positive value to a negative value, the moment is determined to be the moment of direction reversal of the key joint, and the set of direction reversal moments of each key joint is obtained. The reversal times of different key joints in the set of reversal times are clustered in time. Multiple reversal times with a time difference less than a preset time window are grouped into the same cluster, and the center time of each cluster is selected as the attitude transition point.

[0031] like Figure 2As shown, the method includes: After acquiring the multimodal data sequence, the first step is to select key joints highly correlated with the shooting motion from the joint position sequence. These key joints include the wrist, elbow, shoulder, and knee joints, which play crucial motor roles in the shooting motion. The correlation between each joint and the shooting motion is calculated, specifically by comprehensively evaluating factors such as the displacement amplitude and velocity change characteristics of each joint during the shooting process in a large amount of labeled dataset. For example, in a certain test dataset, the correlation of the wrist joint is 0.92, the elbow joint is 0.89, the shoulder joint is 0.85, and the knee joint is 0.78, all exceeding the preset correlation threshold of 0.75. Therefore, these four joints are selected as key joints. The three-dimensional spatial coordinate data of these key joints are extracted from the original joint position sequence to form a key joint position subsequence.

[0032] For the extracted key joint position subsequence, the velocity vector of each key joint at consecutive time points is calculated. Specifically, this is done by performing a difference operation on the position coordinates of each key joint at adjacent time points. Assuming the position coordinates of a key joint at time t are (x_t, y_t, z_t), and the position coordinates at time t+1 are (x_{t+1}, y_{t+1}, z_{t+1}), then the velocity vector of this joint at time t+1 can be expressed as (x_{t+1}-x_t, y_{t+1}-y_t, z_{t+1}-z_t). This difference calculation is performed on each key joint at each sampling time point to obtain the velocity vector sequence of each key joint. For example, in a basketball shooting motion capture, data is collected at a frequency of 30 frames per second. The position of the wrist joint in the 10th frame is (0.45, 1.20, 0.30) meters, and the position in the 11th frame is (0.48, 1.25, 0.32) meters. Then the velocity vector in the 11th frame is calculated to be (0.03, 0.05, 0.02) meters / frame.

[0033] After obtaining the velocity vector sequence, the dot product operation is performed on the velocity vectors of each key joint at adjacent time points to determine the change in the direction of joint movement. The dot product operation effectively reflects the directional relationship between two vectors: a positive value indicates that the directions are basically the same, and a negative value indicates that the direction deviates significantly. The dot product value of the velocity vector at each time point with the velocity vector at the previous time point is calculated. When the dot product value changes from a positive value to a negative value, it indicates that the direction of movement of the joint has reversed. For example, the velocity vector of the wrist joint in frame 25 is (0.06, 0.08, 0.04), and the velocity vector in frame 26 is (-0.03, -0.02, -0.01). The dot product of the velocity vectors in the two frames is calculated as 0.06×(-0.03)+0.08×(-0.02)+0.04×(-0.01)=-0.0036. Since the result is negative and the dot product of the previous time point is positive, frame 26 is determined to be a time when the direction of the wrist joint reverses. In this way, the reversal times of all key joints are identified, forming a set of reversal times.

[0034] For the identified set of reversal moments, further temporal clustering is performed. Multiple joints with similar reversal moments are grouped together to identify the posture transition points in the overall movement. A preset time window is set, such as 5 frames (approximately 0.17 seconds). When the time difference between the reversal moments of different key joints is less than this window value, they are considered to represent the same posture transition stage and are grouped into the same cluster. For each cluster, its center moment is selected as the posture transition point. The center moment can be obtained by calculating the average of all reversal moments within that cluster. For example, in the analysis of a basketball shooting motion, reversal was detected in the shoulder joint at frame 42, the elbow joint at frame 45, and the wrist joint at frame 44. The time difference between these three moments is less than the 5-frame time window, so they are grouped into the same cluster. The average is calculated, and frame 43.67 (rounded down to frame 44) is taken as a posture transition point.

[0035] Through the above processing, key posture transition points can be accurately identified from complex shooting motion sequences. These transition points typically correspond to key technical aspects of the shooting motion, such as knee flexion for power, arm extension, and wrist rotation. In practical applications, parameters such as the correlation threshold and time window need to be dynamically adjusted based on different users' physical characteristics and exercise habits to improve the accuracy and adaptability of posture transition point identification. Furthermore, combining other sensor data, such as pressure and acceleration, can further enhance the accuracy of identification.

[0036] In one optional implementation, a stage completion detection is performed on the segmented data of each stage. When it is detected that the segmented data of a certain stage meets the completion judgment condition, an evaluation trigger signal for that stage is generated, including: Real-time duration statistics are performed on the segmented data of each stage. By recording the start time and current time of the segmented data of each stage, the cumulative duration of the segmented data of each stage is calculated. The cumulative duration is compared with the preset minimum duration threshold corresponding to the stage. When the cumulative duration reaches the preset minimum duration threshold, a duration marker for the stage is generated. The characteristic fluctuation degree of segmented data in each stage within a continuous time window is monitored, and the variance index of the characteristic fluctuation degree is calculated. When the variance index is lower than the preset stability threshold and remains below the preset stability duration, the stage is determined to have entered a stable state and a stability achievement mark for the stage is generated. When a certain stage simultaneously possesses both the duration target identifier and the stability target identifier, it is determined that the segmented data of that stage meets the completion determination condition, and an evaluation trigger signal containing the stage identifier and completion time information is generated.

[0037] First, the system receives input signals containing data from multiple stages. These signals can originate from industrial process monitoring sensors, medical equipment physiological signal acquisition systems, or intelligent driving status monitoring devices. After preprocessing, the input signals are divided into stages, segmenting the continuous data stream into multiple data segments based on preset characteristic boundary conditions. For example, in a production line monitoring scenario, the data can be divided according to process stages such as raw material input, heating, cooling, and forming; in a medical monitoring scenario, it can be divided into physiological state stages such as rest, exercise, and recovery.

[0038] For the segmented data of each stage, real-time duration statistics are performed. The duration statistics module records the start time (timestamp_start) and current time (timestamp_current) of each stage, and calculates the cumulative duration (duration = timestamp_current - timestamp_start). The calculated duration is compared with the preset minimum duration threshold (min_duration_threshold) for that stage. Taking medical monitoring as an example, if the minimum duration threshold for the exercise stage is set to 300 seconds, when the actual exercise duration reaches or exceeds 300 seconds, a duration flag (duration_flag = true) is generated for that stage. If the threshold is not reached, duration_flag = false. This flag is recorded in the stage status table as one of the necessary conditions for determining whether the stage is completed.

[0039] Simultaneously, the characteristic fluctuation degree of segmented data in each stage within a continuous time window is monitored. The fluctuation monitoring module first defines a sliding time window `window_size`, such as 30 seconds, and collects data points within this window. For the data point set `data_window` within the current time window, the variance index of its characteristic fluctuation degree is calculated. Taking industrial production temperature monitoring as an example, the variance of temperature data points collected in the most recent 30 seconds of the heating stage is calculated, resulting in `variance_index = 0.85`. This variance index is compared with a preset stability threshold `stability_threshold` (such as 1.2). When `variance_index < stability_threshold`, the start time of the current stable state `stable_start_time` is recorded. Monitoring continues, and when the duration of the stable state `stable_duration = current_time - stable_start_time` exceeds the preset stable duration `min_stability_duration` (such as 60 seconds), it is determined that the stage has entered a stable state, and a stability flag `stability_flag = true` is generated for this stage. If the characteristic fluctuation exceeds the stability threshold or the stable duration is insufficient, then stability_flag = false.

[0040] The completion determination module comprehensively evaluates the duration flag and the stability flag. When a certain stage simultaneously meets both duration_flag = true and stability_flag = true, the segmented data of that stage is determined to meet the completion determination conditions. For example, in an intelligent driving scenario, when the cruise phase duration reaches 600 seconds (the preset minimum duration threshold), and the vehicle speed fluctuation variance is 0.5 (below the preset stability threshold of 0.8) and remains stable for more than 120 seconds (the preset stable duration), the cruise phase is determined to be completed, and an evaluation trigger signal is generated.

[0041] The evaluation trigger signal generation module creates an evaluation trigger signal `trigger_signal` containing a phase identifier `phase_id` and a completion time `completion_time`. The data structure of this signal is `{phase_id: "cruise", completion_time: "2023-05-18 10:45:32", duration: 642, stability_value: 0.5}`, where `phase_id` identifies the phase name, `completion_time` records the precise timestamp of completion determination, `duration` records the actual duration, and `stability_value` records the final stability index value. This evaluation trigger signal is then passed to subsequent processing modules to initiate the in-depth analysis and evaluation process of the phase data.

[0042] A dynamic adjustment mechanism is set up to adaptively adjust the preset minimum duration threshold and preset stability threshold for each stage based on historical data statistics. For example, by analyzing data from the past 30 production processes, it was found that the average duration of the cooling stage is 450 seconds and the standard deviation is 45 seconds. Therefore, the minimum duration threshold for the cooling stage is adjusted from 400 seconds to 405 seconds. At the same time, based on the historical distribution of fluctuations in the steady-state characteristics, the preset stability threshold is adjusted from 1.0 to 0.95 to improve the accuracy of the judgment.

[0043] This method effectively solves the problem of inaccurate judgment by a single indicator through a dual judgment mechanism of duration statistics and fluctuation stability, improves the reliability of stage completion detection, and provides a solid foundation for subsequent accurate evaluation and intelligent decision-making.

[0044] In one optional implementation, in response to the evaluation trigger signal, a posture score is calculated on the motion feature vector of the corresponding stage to obtain the posture score for that stage, including: In response to the evaluation trigger signal, when the stage corresponding to the evaluation trigger signal is the force exertion stage or the follow-up stage, the posture score of the preceding stage of the shooting action and the action feature vector of the preceding stage are extracted from the buffer. When the stage corresponding to the evaluation trigger signal is the preparation stage, the stage is marked as the starting stage and the preceding data extraction is skipped. For non-initial stages, a transition smoothness index is calculated between the action feature vector of the corresponding stage and the action feature vector of the preceding stage. The transition smoothness index is obtained by calculating the continuity measure of the two feature vectors in a preset transition feature space. A coherence correction coefficient is generated based on the transition smoothness index. The motion feature vector of the corresponding stage is compared with the standard posture template of the corresponding stage, the independent posture deviation value of the stage is calculated, and the independent score value of the stage is generated based on the independent posture deviation value. For non-initial stages, the independent score value is multiplied by the coherence correction coefficient and then superimposed with the inherited correction amount calculated based on the attitude score of the preceding stage to obtain the attitude score for that stage. For the initial stage, the independent score value is directly used as the attitude score for that stage.

[0045] When an evaluation trigger signal is detected, it is determined which shooting phase the signal corresponds to. If the evaluation trigger signal corresponds to the preparation phase, this phase is marked as the starting phase. At this time, no preceding data needs to be extracted because the preparation phase is the first phase of the shooting motion. If the evaluation trigger signal corresponds to the power generation phase or the follow-up phase, the posture score and motion feature vector of the preceding phase of the shooting motion are extracted from the buffer. For example, when evaluating the power generation phase, the posture score (e.g., 85 points) and motion feature vector of the preparation phase (e.g., [0.78, 0.92, 0.85, 0.88], representing key joint angles and position information) of the preparation phase are extracted.

[0046] For non-initiation phases (i.e., the power generation and follow-up phases), a smoothness index is calculated between the motion feature vector of the current phase and the motion feature vector of the preceding phase. This index is obtained by calculating the continuity measure of the two feature vectors in a preset transition feature space. Specifically, the rate of change of corresponding elements in the feature vectors of the two phases is analyzed to calculate the smoothness of joint angle changes. For example, if the motion feature vector of the power generation phase is [0.65, 0.87, 0.92, 0.79], its smoothness of transition with the feature vector of the preparation phase [0.78, 0.92, 0.85, 0.88] is calculated. By calculating the consistency and smoothness of the changes in each component, a smoothness index value of 0.92 is obtained (1 for complete smoothness). Based on this index, a coherence correction coefficient of 0.95 is generated for subsequent scoring adjustments.

[0047] While acquiring the transition smoothness index, the motion feature vector of the current stage is compared with the corresponding standard posture template for that stage. The standard posture template is a pre-stored ideal motion feature vector; for example, the standard template for the force exertion stage is [0.70, 0.90, 0.95, 0.80]. The difference between the motion feature vector of the current stage and the standard template is calculated to obtain the independent posture deviation value. If the feature vector of the force exertion stage is [0.65, 0.87, 0.92, 0.79], the independent posture deviation value calculated after comparing it with the standard template is 0.06. According to the preset scoring mapping rule, this deviation value is converted into an independent score of 88 points.

[0048] For the non-initial stage, the independent score is further adjusted by multiplying it by a consistency correction coefficient. In the example of the power exertion stage, 88 points multiplied by 0.95 yields 83.6 points. Simultaneously, an inheritance correction is calculated based on the attitude scores of the preceding stages. Assuming a 20% inheritance weight, the inheritance correction is 20% of the preparation stage score of 85 points, or 17 points. Finally, the attitude score for the power exertion stage is 83.6 points plus 17 points, equaling 100.6 points. Considering the scoring cap of 100 points, the final score for the power exertion stage is output as 100 points.

[0049] For the initial stage (preparation stage), since there is no preceding stage, the independent score is directly used as the posture score for this stage. For example, if the motion feature vector of the preparation stage is [0.78, 0.92, 0.85, 0.88], and the independent posture deviation value after comparison with the standard template of the preparation stage [0.80, 0.95, 0.85, 0.90] is 0.03, the corresponding independent score is 85 points, then the final posture score of the preparation stage is 85 points.

[0050] In practical implementation, different transition feature weights can be set to emphasize the smooth transition of certain key joints. For example, the continuity of movement in the wrist and elbow is more important than that in the foot, so the corresponding feature components will have higher weights. Similarly, when calculating independent posture deviations, different weights are applied to different joint angles. For example, when shooting, the wrist angle has a weight of 0.4, the elbow 0.3, the shoulder 0.2, and the leg 0.1, to reflect the degree of influence of different body parts on shooting accuracy.

[0051] This method not only assesses the standard of posture at each stage but also considers the fluidity between movements, which is particularly important for assessing sports skills requiring high coordination, such as shooting. The scoring results can be provided to users in real time, helping them understand their strengths and weaknesses and thus enabling targeted training and improvement.

[0052] In one optional implementation, a transition smoothness index is calculated between the action feature vector of the corresponding stage and the action feature vector of the preceding stage. The transition smoothness index is obtained by calculating the continuity measure of the two feature vectors in a preset transition feature space. Generating a coherence correction coefficient based on the transition smoothness index includes: The action feature vectors of the corresponding stage and the action feature vectors of the preceding stage are projected onto the preset transition feature space to obtain the projected feature vectors of the corresponding stage and the projected feature vectors of the preceding stage. Extract the end feature component of the projection feature vector of the corresponding stage in the preset transition feature space and the start feature component of the projection feature vector of the preceding stage in the preset transition feature space, calculate the feature distance between the end feature component and the start feature component, and construct the stage connection deviation based on the feature distance; Perform time-series derivative operations on the projected feature vectors of the corresponding stage and the projected feature vectors of the preceding stage to obtain the feature rate of change vectors of the corresponding stage and the preceding stage. Calculate the directional consistency metric between the feature rate of change vectors of the corresponding stage and the preceding stage, and construct a trend continuity index based on the directional consistency metric. The transition smoothness index is obtained by fusing the stage connection deviation and the trend continuity index; the transition smoothness index is then mapped to a continuity correction coefficient with a value range within a preset interval through coherence mapping.

[0053] To calculate the transition smoothness index, the action feature vectors of the corresponding stage and the preceding stage need to be projected onto a preset transition feature space. Specifically, assuming the action feature vector of the current stage is [0.3, 0.5, 0.7, 0.2] and the action feature vector of the preceding stage is [0.1, 0.4, 0.6, 0.3], these two vectors are projected onto the preset transition feature space using a feature projection matrix. This projection matrix can be a matrix that has been pre-trained using a large amount of action data and is capable of extracting features related to action coherence. After projection, the projected feature vector of the corresponding stage is [0.25, 0.45], and the projected feature vector of the preceding stage is [0.15, 0.35].

[0054] After projection, key feature components are extracted from the two projected feature vectors. The final feature component is extracted from the projected feature vector of the corresponding stage, for example, the feature value of the last time point [0.45] is taken; the initial feature component is extracted from the projected feature vector of the preceding stage, for example, the feature value of the first time point [0.15] is taken. The feature distance between these two feature components is calculated using the Euclidean distance method, resulting in a feature distance of 0.3. This feature distance represents the feature difference between the two stages at the junction; a larger distance indicates a higher degree of abrupt change at the junction. The stage junction deviation is constructed based on the feature distance and can be normalized by setting a threshold. For example, setting the threshold to 1.0 results in a normalized stage junction deviation of 0.3.

[0055] In addition to calculating the distance between feature components, it is also necessary to assess the continuity of the trend of action change. This involves performing a temporal derivative operation on the projected feature vector to obtain the feature rate of change vector. Specifically, the derivative can be approximated by calculating the difference between feature values ​​at adjacent time points. Processing the aforementioned projected feature vector yields a feature rate of change vector of [0.2] for the corresponding stage and [0.2] for the preceding stage. The directional consistency measure between these two feature rate of change vectors is calculated using the cosine of the angle between the vectors, resulting in a value of 1.0, indicating that the trends of change in the two stages are completely consistent. Based on this consistency measure, a trend continuity index is constructed. For example, the cosine value can be directly used as the continuity index, or it can be processed using a nonlinear mapping, resulting in a trend continuity index of 1.0.

[0056] The transition smoothness index is obtained by merging the stage transition deviation and the continuity index of the change trend. The fusion method can be a weighted average. For example, if the weight of the transition deviation is set to 0.4 and the weight of the continuity index is set to 0.6, then the transition smoothness index is calculated as: 0.4×(1-0.3)+0.6×1.0=0.88. The higher the index value, the smoother the transition.

[0057] Finally, the transition smoothness index is mapped to a consistency correction coefficient using a consistency mapping function. This mapping function can be designed as a piecewise linear function or a sigmoid-like function, limiting the range of the correction coefficient to a preset interval, such as [0.5, 1.5]. In this example, assuming the mapping function is a linear mapping with range adjustment: correction coefficient = 0.5 + transition smoothness index × 1.0, then the consistency correction coefficient is 0.5 + 0.88 × 1.0 = 1.38.

[0058] In practical applications, the parameters can be adjusted based on the specific needs of the scenario. For example, in fast-paced action scenarios, the weight of the stage transition deviation can be increased to emphasize the accuracy of the movements; for artistic performances such as dance, the weight of the continuity index of the changing trend can be increased to emphasize the smoothness of the movements. Through this flexible parameter adjustment, this method can adapt to the needs of various motion generation scenarios.

[0059] This coherence correction coefficient can be used to adjust key parameters in the motion generation process, such as motion amplitude, speed, or transition duration, so that the generated motion sequence is more natural and smooth at the stage transitions, avoiding obvious abrupt changes or inconsistencies.

[0060] This invention relates to a low-power real-time basketball motion recognition system based on artificial intelligence, the system comprising: The first unit is used to collect motion data and visual data during the basketball shooting process through a multimodal sensor. The motion data includes acceleration signals and angular velocity signals, and the visual data includes joint position sequences. The second unit is used to perform time alignment processing on the motion data and the visual data to obtain a time-synchronized multimodal data sequence; The third unit is used to decompose the multimodal data sequence based on the kinematic constraint model, decomposing the continuous shooting action into a preparation phase, a power generation phase, and a follow-up phase, and obtaining segmented data for each phase. The fourth unit is used to extract action feature vectors from the segmented data of each stage. The fifth unit is used to perform stage completion detection on the segmented data of each stage. When it is detected that the segmented data of a certain stage meets the completion judgment condition, an evaluation trigger signal is generated for that stage. The completion judgment condition includes the duration of the stage reaching a preset lower limit and the change in the key feature of the stage reaching a preset change range. The sixth unit is used to calculate the posture score of the action feature vector of the corresponding stage in response to the evaluation trigger signal, and obtain the posture score of the stage. After completing the posture score calculation of all stages, the posture scores of each stage are weighted and fused to obtain the comprehensive posture score. The seventh unit is used to output the comprehensive posture score as the shooting posture evaluation result.

[0061] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0062] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0063] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A low-power real-time basketball action recognition method based on artificial intelligence, characterized in that, The method comprises the following steps: Collecting motion data and visual data during a basketball shooting process through a multi-modal sensor, the motion data including acceleration signals and angular velocity signals, and the visual data including joint position sequences; Time aligning the motion data and the visual data to obtain time-synchronized multi-modal data sequences; Decomposing the multi-modal data sequences based on a kinematic constraint model to divide continuous shooting actions into a preparation phase, a force phase, and a following phase, and obtaining segmented data of each phase; Extracting action feature vectors from the segmented data of each phase respectively; Performing phase completion degree detection on the segmented data of each phase, generating an evaluation trigger signal for a phase when the segmented data of the phase meets a completion degree judgment condition, and the completion degree judgment condition including that the duration of the phase reaches a preset lower limit and the key feature variation of the phase reaches a preset variation amplitude; In response to the evaluation trigger signal, performing posture score calculation on the action feature vector of the corresponding phase to obtain the posture score of the phase, and after completing the posture score calculation of all phases, weighting and fusing the posture scores of each phase to obtain a comprehensive posture score; Outputting the comprehensive posture score as a shooting posture evaluation result.

2. The method of claim 1, wherein, The method comprises the following steps: Performing peak detection on the acceleration signals in the multi-modal data sequences to identify local maximum points and local minimum points of acceleration amplitude as motion intensity transition points, and performing trajectory analysis on the joint position sequences in the multi-modal data sequences to identify time points at which the joint motion direction reverses as posture transition points, to obtain a candidate boundary point set containing motion intensity transition points and posture transition points; Projecting the motion intensity transition points and the posture transition points in the candidate boundary point set to a unified time axis, calculating the matching degrees of each candidate boundary point with phase boundary features in a preset kinematic template, selecting candidate boundary points with matching degrees exceeding a preset matching threshold as phase boundary points, and the phase boundary points including a first boundary point between the preparation phase and the force phase and a second boundary point between the force phase and the following phase; Dividing the multi-modal data sequences into first segmented data corresponding to the preparation phase, second segmented data corresponding to the force phase, and third segmented data corresponding to the following phase, to obtain segmented data of each phase, with the first boundary point and the second boundary point as segmentation time points.

3. The method of claim 2, wherein, The method comprises the following steps: Selecting key joints with an association degree exceeding a preset association threshold from the joint position sequences, the key joints including wrist joints, elbow joints, shoulder joints, and knee joints, to obtain key joint position subsequences; Differential operation is performed on position coordinates of each key joint in the key joint position subsequence at continuous time points, motion velocity vectors of each key joint at each time point are calculated, and a velocity vector sequence of each key joint is obtained; Inner product operation is performed on velocity vectors of each key joint in the velocity vector sequence at adjacent time points, when the inner product value of the velocity vector of a certain key joint at a time point and the velocity vector of the key joint at a previous time point changes from a positive value to a negative value, it is determined that the time point is a direction reverse time point of the key joint, and a direction reverse time point set of each key joint is obtained; The direction reverse time points of different key joints in the direction reverse time point set are time clustered, a plurality of direction reverse time points with a time difference less than a preset time window are classified into a same clustering cluster, and a center time point of each clustering cluster is selected as the posture conversion point.

4. The method of claim 1, wherein, The stage completion degree detection is performed on the segmented data of each stage, and when it is detected that the segmented data of a certain stage meets the completion degree determination condition, an evaluation trigger signal for the stage is generated, including: The real-time duration of the segmented data of each stage is counted, the starting time point and the current time point of each stage segmented data are recorded, the cumulative duration of each stage segmented data is calculated, the cumulative duration is compared with the preset minimum duration threshold corresponding to the stage, and when the cumulative duration reaches the preset minimum duration threshold, a duration up-to-standard identifier of the stage is generated; The feature fluctuation degree of the segmented data of each stage within a continuous time window is monitored, a variance index of the feature fluctuation degree is calculated, and when the variance index is lower than a preset stability threshold and remains for more than a preset stability duration, it is determined that the stage enters a stable state and a stable up-to-standard identifier of the stage is generated; When a stage has both the duration up-to-standard identifier and the stable up-to-standard identifier, it is determined that the segmented data of the stage meets the completion degree determination condition, and an evaluation trigger signal containing the stage identifier and completion time information is generated.

5. The method of claim 1, wherein, In response to the evaluation trigger signal, posture score calculation is performed on the action feature vector of the corresponding stage, and the posture score of the stage is obtained, including: In response to the evaluation trigger signal, when the stage corresponding to the evaluation trigger signal is a force stage or a follow-up stage, the posture score of the previous stage of the shot motion sequence and the action feature vector of the previous stage are extracted from the cache area, and when the stage corresponding to the evaluation trigger signal is a preparation stage, the stage is marked as a starting stage and the previous data extraction is skipped; For non-starting stages, a transition smoothness index between the action feature vector of the corresponding stage and the action feature vector of the previous stage is calculated, the transition smoothness index is obtained by calculating the continuity measurement value of the two feature vectors in a preset transition feature space, a continuity correction coefficient is generated based on the transition smoothness index; The action feature vector of the corresponding stage is compared with the standard posture template corresponding to the stage, an independent posture deviation value of the stage is calculated, and an independent score value of the stage is generated based on the independent posture deviation value. For the non-starting stage, the independent score value is multiplied by the continuity correction coefficient, and an inherited correction amount calculated based on the posture score of the previous stage is superimposed to obtain the posture score of the stage; for the starting stage, the independent score value is directly taken as the posture score of the stage.

6. The method of claim 5, wherein, A transition smoothness index between the action feature vector of the corresponding stage and the action feature vector of the previous stage is calculated, the transition smoothness index being obtained by calculating a continuity measure value of the two feature vectors in a preset transition feature space, and the continuity correction coefficient is generated based on the transition smoothness index, including: The action feature vector of the corresponding stage and the action feature vector of the previous stage are projected into the preset transition feature space respectively to obtain a projected feature vector of the corresponding stage and a projected feature vector of the previous stage; The end feature component of the projected feature vector of the corresponding stage in the preset transition feature space and the start feature component of the projected feature vector of the previous stage in the preset transition feature space are extracted, a feature distance between the end feature component and the start feature component is calculated, and a stage connection deviation amount is constructed based on the feature distance; The projected feature vector of the corresponding stage and the projected feature vector of the previous stage are subjected to time derivative operation to obtain a feature change rate vector of the corresponding stage and a feature change rate vector of the previous stage, a directional consistency measure value between the feature change rate vector of the corresponding stage and the feature change rate vector of the previous stage is calculated, and a change trend continuity index is constructed based on the directional consistency measure value; The stage connection deviation amount and the change trend continuity index are fused to obtain the transition smoothness index; and the transition smoothness index is mapped into the continuity correction coefficient with a value range in a preset interval through continuity mapping.

7. A low-power real-time basketball action recognition system based on artificial intelligence for implementing the method according to any one of claims 1 to 6, characterized in that, It includes: The first unit is configured to collect motion data and visual data in a basketball shooting process through a multi-modal sensor, the motion data including acceleration signals and angular velocity signals, and the visual data including joint position sequences. The second unit is configured to perform time alignment processing on the motion data and the visual data to obtain a time-synchronized multi-modal data sequence. The third unit is configured to perform decomposition processing on the multi-modal data sequence based on a kinematic constraint model, decompose a continuous shooting action into a preparation stage, a force stage and a following stage, and obtain segmented data of each stage. The fourth unit is configured to extract an action feature vector from the segmented data of each stage. The fifth unit is configured to perform stage completion degree detection on the segmented data of each stage, generate an evaluation trigger signal for a corresponding stage when it is detected that the segmented data of the stage meets a completion degree judgment condition, and the completion degree judgment condition includes that the duration of the stage reaches a preset lower limit of time length and the key feature change amount of the stage reaches a preset change amplitude. The sixth unit is configured to perform posture score calculation on the action feature vector of the corresponding stage in response to the evaluation trigger signal to obtain the posture score of the stage, and to obtain a comprehensive posture score by weighted fusion of the posture scores of the stages after completing the posture score calculation of all stages. A seventh unit configured to output the comprehensive shooting posture score as a shooting posture evaluation result.

8. An electronic device, comprising: Comprise: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to perform the method of any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions, when executed by the processor, implement the method of any one of claims 1 to 6.

Images