Motion function evaluation method and system based on deep learning

Abstract

The invention relates to a motion function evaluation method and system based on deep learning, and the method comprises the steps: obtaining a motion video of a subject, detecting a joint key point through employing a posture estimation algorithm, dynamically adjusting the weight based on the confidence degree to optimize the detection stability of a lower limb joint, and carrying out the standardization processing of a coordinate, and obtaining a time sequence; a bidirectional recurrent neural network is used for extracting time sequence features, target joint feature weights in a key stage are enhanced through a phase perception attention mechanism, kinematics differential modeling is carried out on a specific joint track, and the specific joint track is coded into a fixed-dimension feature vector; finally, general biomechanical characteristics are extracted through a parameter sharing mechanism in a multi-task analysis network, and a quantitative evaluation result containing the joint motion range, the muscle compensation mode and the kinematic chain function is generated in combination with task specificity analysis and biomechanical constraint rules. According to the method, the problems of high subjectivity and low efficiency of a traditional evaluation method are solved, and automatic, high-precision and low-cost exercise function evaluation is realized.

Description

Technical Field

[0001] This invention relates to the field of motor function assessment technology, and in particular to a motor function assessment method and system based on deep learning. Background Technology

[0002] Functional Movement Screening (FMS) is a widely used assessment method in rehabilitation medicine, sports science, and physical training. It aims to systematically detect abnormalities in the human kinetic chain, limitations in joint range of motion, and muscle compensation patterns. This type of assessment is of great significance for sports injury prevention, evaluation of rehabilitation treatment effectiveness, and optimization of athletic performance.

[0003] Currently, common methods for assessing motor function mainly fall into two categories: one is the traditional FMS based on manual observation, where physical therapists or coaches observe subjects performing standard movements (such as overhead squats, single-leg squats, etc.) with the naked eye and then give subjective scores based on experience; the other is the objective assessment method based on high-precision instruments, such as wearable sensors like inertial measurement units (IMUs) or optical motion capture systems (such as Vicon, Qualisys, etc.) to collect motion data, which is then analyzed using specialized software.

[0004] Existing methods for assessing motor function suffer from the following problems: Firstly, manual observation assessment methods are highly subjective, easily influenced by the assessor's experience, and lack consistency. The assessment process is time-consuming, typically requiring 20-30 minutes per person, resulting in low efficiency. Furthermore, the human eye struggles to capture subtle compensatory movements occurring during high-speed or complex movements (such as momentary knee valgus or abnormal scapular glide), leading to missed or incorrect diagnoses. Secondly, while wearable devices or optical capture systems can provide quantitative data, the equipment is expensive and requires specialized operation and environmental support, hindering widespread application. Moreover, existing technologies often focus on single movements or local joint analysis, lacking the ability to comprehensively assess multi-movement and multi-joint coordination patterns, thus failing to fully reflect kinetic chain dysfunction. Therefore, developing a low-cost, automated motor function assessment system capable of multi-task collaborative analysis is of great significance. Summary of the Invention

[0005] In view of this, it is necessary to provide a deep learning-based method and system for evaluating motion function to solve the above-mentioned problems in the prior art.

[0006] To address the aforementioned problems, in a first aspect, embodiments of the present invention provide a deep learning-based method for evaluating motor function, comprising:

[0007] The study acquires motion videos of subjects performing functional movements, uses a pose estimation algorithm to detect human joint key points in each image frame of the motion video, dynamically adjusts joint weights based on joint detection confidence to optimize the stability of lower limb joint detection, and standardizes the original coordinates of joint key points in each frame to obtain a time series of standardized joint key point coordinates.

[0008] Based on the time series of the standardized joint key point coordinates, a bidirectional recurrent neural network is used to extract the action temporal features. The target joint feature weights are enhanced in the key motion phase through a phase-aware attention mechanism. Kinematic differential modeling is performed on the motion trajectory of a specific joint, and the temporal features are encoded into a fixed-dimensional feature vector.

[0009] The feature vectors are input into a multi-task analysis network, and common biomechanical features across different functional movements are extracted through a parameter sharing mechanism. At the same time, biomechanical characteristic analysis is performed on each functional movement, and quantitative evaluation results including joint range of motion, muscle compensation patterns, and kinetic chain function are generated by combining predefined biomechanical constraint rules.

[0010] Preferably, the step of using a pose estimation algorithm to detect key points of human joints in each image frame of a motion video, and dynamically adjusting joint weights based on joint detection confidence to optimize the stability of lower limb joint detection, specifically includes:

[0011] Obtain the initial confidence level C for each joint. j The weights of each joint are dynamically calculated using the Sigmoid function; the weight calculation formula is as follows:

[0012]

[0013] In the formula, W j C is the adjusted weight of joint j. j Let θ be the initial confidence level of joint j, k be the slope, and θ be the confidence threshold.

[0014] Based on the aforementioned weights, priority is given to improving the detection stability of lower limb joints, ensuring the reliability of the detection results of key points of lower limb joints.

[0015] Preferably, the standardization process for the original coordinates of joint key points in each frame to obtain a time series of standardized joint key point coordinates specifically includes:

[0016] Using the center point of the pelvis as the origin of the local coordinate system, obtain the original coordinates of the center point of the pelvis;

[0017] The original coordinates of joint keypoints in each frame are standardized using the Euclidean distance between the two shoulder joints as the lateral dimension reference and the vertical distance from the top of the head to the ankle joint as the longitudinal dimension reference. The standardization formula is as follows:

[0018]

[0019] In the formula, x and y are the original coordinates of the joint key points, and x' and y' are the standardized coordinates of the joint key points; x pelvis and y pelvis represents the original x and y coordinates of the center point of the pelvis, shoulder_width is the Euclidean distance between the two shoulder joints, and body_height is the vertical distance from the top of the head to the ankle joint.

[0020] A sliding window normalization process is performed on the standardized joint coordinates of multiple consecutive frames to form a time series of standardized joint key point coordinates.

[0021] Preferably, the step of extracting motion temporal features using a bidirectional recurrent neural network based on the time series of the standardized joint key point coordinates, and enhancing the target joint feature weights during key motion phases through a phase-aware attention mechanism, specifically includes:

[0022] A bidirectional long short-term memory network is used to process the time series of standardized joint key point coordinates. Forward LSTM captures the forward temporal dependency, and backward LSTM captures the backward temporal dependency.

[0023] A phase-aware attention mechanism is introduced to identify key motion phases of an action through kinematic features; the attention weights of the key motion phases are calculated based on the hidden states output by the bidirectional long short-term memory network, using the following formula:

[0024] Attention Weight = Sigmoid(W) a ·h t +b a )

[0025] In the formula, W a and b a h are learnable parameters, AttentionWeight is the attention weight, and h is the attention weight. t The hidden state at time step t;

[0026] The attention weights enhance the feature weights of target joints in key motion phases, thereby strengthening the expression of target joint features in temporal features.

[0027] Preferably, the step of performing kinematic differential modeling on the motion trajectory of a specific joint and encoding the temporal features into a fixed-dimensional feature vector specifically includes:

[0028] Second-order kinematic differential modeling is performed on the scapular motion trajectory to calculate joint motion velocity and acceleration;

[0029] The temporal features enhanced with target joint weights are fused with the kinematic differential modeling results of specific joints, and then input into a fully connected layer for feature compression, encoding them into fixed-dimensional feature vectors.

[0030] Preferably, the step of inputting the feature vector into a multi-task analysis network and extracting common biomechanical features across different functional actions through a parameter sharing mechanism specifically includes:

[0031] The feature vector is input into a multi-task analysis network. Through the network's built-in parameter sharing mechanism, common biomechanical features across different functional actions are extracted, and these common biomechanical features are reused through the parameter sharing mechanism.

[0032] The common biomechanical characteristics include core stability mode, basic joint range of motion, and kinetic chain transmission efficiency. Core stability mode is a characteristic that reflects the coordinated control ability of the trunk and pelvis. Basic joint range of motion is a characteristic that reflects the basic range of motion of the hip and shoulder joints. Kinetic chain transmission efficiency is a characteristic that reflects the effectiveness of force transmission from the lower limbs to the upper limbs.

[0033] Preferably, the biomechanical characteristic analysis for each functional movement, combined with predefined biomechanical constraint rules to generate quantitative assessment results including joint range of motion, muscle compensation patterns, and kinetic chain function, specifically includes:

[0034] In the multi-task analysis network, specific analyses are performed for different functional movements. For the overhead squat, the coordination of the hip-knee-ankle joint chain is analyzed; for the angel on the wall, the stability of the scapular trajectory is analyzed; and for the single-leg squat, the coronal plane control of the pelvis is analyzed.

[0035] The rationality of the specific analysis results is verified by using predefined biomechanical constraint rules. Based on the verified specific analysis results, quantitative assessment results including joint range of motion, muscle compensation patterns and kinetic chain function are generated.

[0036] Secondly, embodiments of the present invention provide a deep learning-based motion function evaluation system, comprising:

[0037] The pose estimation module is used to acquire motion videos of subjects performing functional movements. It uses a pose estimation algorithm to detect human joint key points in each image frame of the motion video. Based on the joint detection confidence, it dynamically adjusts the joint weights to optimize the stability of lower limb joint detection. The original coordinates of the joint key points in each frame are standardized to obtain a time series of standardized joint key point coordinates.

[0038] The spatiotemporal feature encoding module is used to extract action temporal features based on the time series of the standardized joint key point coordinates using a bidirectional recurrent neural network, enhance the target joint feature weights in key motion phases through a phase-aware attention mechanism, perform kinematic differential modeling of the motion trajectory of a specific joint, and encode the temporal features into a fixed-dimensional feature vector.

[0039] The multi-task decision module is used to input the feature vector into the multi-task analysis network, extract common biomechanical features across different functional movements through a parameter sharing mechanism, perform biomechanical characteristic analysis for each functional movement, and generate quantitative assessment results including joint range of motion, muscle compensation patterns, and kinetic chain function by combining predefined biomechanical constraint rules.

[0040] Preferably, the multi-task decision module includes:

[0041] A shared layer is used to extract common biomechanical features of all functional movements, including core stability patterns, basic joint range of motion, and kinetic chain transmission efficiency.

[0042] The exclusive layer, connected to the shared layer, is used for specific analysis of various functional movements, including analyzing the coordination of the hip-knee-ankle joint chain for overhead squats, analyzing the stability of the scapular trajectory for the angel on the wall, and analyzing the coronal plane control of the pelvis for single-leg squats.

[0043] The output layer, connected to the exclusive layer, is used to integrate biomechanical constraint rules and verify the rationality of the analysis results, generating quantitative assessment results including joint range of motion, muscle compensation patterns, and kinetic chain function.

[0044] Thirdly, the present invention also provides an electronic device, including a memory and a processor, wherein,

[0045] The memory is used to store programs;

[0046] The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the deep learning-based motion function evaluation method as described in the first aspect embodiment of the present invention.

[0047] The deep learning-based motion function evaluation method and system provided by this invention have the following advantages compared with the prior art:

[0048] 1) This invention uses computer vision technology and a pose estimation algorithm to detect key points of human joints, and performs automated analysis based on a deep learning model. This avoids the problems of strong subjectivity and poor consistency caused by differences in the experience of evaluators in manual observation and evaluation, making the evaluation results more objective and accurate.

[0049] 2) This invention extracts action temporal features through a bidirectional recurrent neural network, strengthens target joint features in key motion phases through a phase-aware attention mechanism, and performs kinematic differential modeling of specific joint trajectories. This replaces human eye observation and single-dimensional analysis, achieving the effect of accurately capturing subtle compensatory movements, improving evaluation accuracy, and reducing missed and false judgments.

[0050] 3) This invention employs a multi-task analysis network, extracting common biomechanical features across different functional movements through a parameter sharing mechanism. Simultaneously, it performs specific analyses for each movement, achieving a holistic assessment of multi-movement, multi-joint coordination patterns and comprehensively reflecting kinetic chain dysfunction. Furthermore, this system analyzes ordinary motion videos, eliminating the need for expensive wearable devices or specialized optical capture systems, thus reducing assessment costs and facilitating widespread application. Attached Figure Description

[0051] Figure 1 Flowchart of the deep learning-based motion function evaluation method provided by this invention;

[0052] Figure 2 The flowchart for assessing poor posture provided by this invention;

[0053] Figure 3 The flowchart for generating a body health report provided by this invention;

[0054] Figure 4 The structural block diagram of the deep learning-based motion function assessment system provided by this invention;

[0055] Figure 5 The data processing flowchart of the deep learning-based motion function assessment system provided by this invention;

[0056] Figure 6 This is a structural block diagram of the electronic device provided by the present invention. Detailed Implementation

[0057] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0058] In the field of motor function assessment, existing technologies mainly focus on two types of approaches: traditional motor function monitoring (FMS) based on manual observation and assessment relying on high-precision instruments. Research on low-cost automated assessment and multi-motor coordination analysis is significantly insufficient. In traditional methods, assessment based on manual observation is limited by the therapist's experience, resulting in strong subjectivity, poor consistency, and an inability to capture subtle compensatory patterns in high-speed movements. While approaches based on wearable sensors or optical capture systems can provide quantitative data, the equipment is expensive, the operation is complex, and they often focus on single movements or local joint analysis, lacking the ability to comprehensively assess multi-motor and multi-joint coordination patterns and thus failing to fully reflect kinetic chain dysfunction.

[0059] In view of this, the present invention aims to provide a deep learning-based method for assessing motor function. By improving the pose estimation algorithm and adaptive joint weight adjustment mechanism, it enhances the stability of keypoint detection and solves the accuracy problems under occlusion and high-speed motion. It employs a bidirectional recurrent neural network and a phase-aware attention mechanism to strengthen the feature representation of key motion phases, achieving accurate capture of subtle compensatory movements. A multi-task analysis network and parameter sharing mechanism are constructed to extract common features across movements and complete specialized analysis. Combined with biomechanical constraint rules, it outputs quantitative assessment results, ultimately achieving low-cost, automated, and multi-task collaborative accurate assessment of motor function. The following will elaborate and describe this method through several embodiments.

[0060] Figure 1 This is a flowchart illustrating the deep learning-based motion function assessment method provided by the present invention. (Refer to...) Figure 1 The deep learning-based motion function evaluation method provided by this invention includes at least the following steps:

[0061] Step S1: Obtain motion videos of the subject performing functional movements, use a pose estimation algorithm to detect human joint key points in each image frame of the motion video, dynamically adjust joint weights based on joint detection confidence to optimize the stability of lower limb joint detection; standardize the original coordinates of joint key points in each frame to obtain a time series of standardized joint key point coordinates.

[0062] Specifically, step S1 is the core step in the data input and preprocessing for motor function assessment in this invention. First, motion videos of the subject performing functional movements are acquired. These movements include overhead squats, wall angels, and single-leg squats. Data acquisition can be achieved using a standard single camera, eliminating the need for expensive wearable devices or professional optical capture systems, thus reducing costs and facilitating widespread application.

[0063] Next, a pose estimation algorithm is used to detect key points of human joints in the video frame by frame. These key points include the ankle, knee, hip, shoulder, pelvis, and top of the head. The detection confidence C for different joints is then calculated. jDifferences exist, and lower limb joints (ankle, knee, hip) are crucial for motor function assessment. Therefore, this embodiment obtains the initial confidence level C for each joint. j The weights of each joint are dynamically calculated using the Sigmoid function. The weight calculation formula is as follows:

[0064]

[0065] In the formula, W j C is the adjusted weight of joint j. j Let θ be the initial confidence level of joint j, k be the slope, and θ be the confidence threshold.

[0066] Joints with higher confidence levels have greater weight, prioritizing the improvement of detection stability for lower limb joints and ensuring the reliability of detection results for key points in lower limb joints. This embodiment prioritizes lower limb joint detection stability through dynamic weighting of confidence levels, avoiding subsequent evaluation deviations due to joint detection errors and solving the problem of difficulty in accurately locating joints through manual observation.

[0067] Furthermore, the original coordinates of the joint key points in each frame are standardized to obtain a time series of standardized joint key point coordinates, including: obtaining the original coordinates of the pelvic center point with the pelvic center point as the origin of the local coordinate system. Since the pelvis is the core hub of the human kinematic chain, using it as the origin can establish a unified human relative coordinate benchmark, avoiding coordinate offsets caused by shooting angles.

[0068] The original coordinates of joint keypoints in each frame were standardized using the Euclidean distance between the two shoulder joints as the lateral dimension reference and the vertical distance from the top of the head to the ankle joint as the longitudinal dimension reference. This standardization eliminates the incomparability of coordinates caused by individual body shape differences. The standardization formula is as follows:

[0069]

[0070] In the formula, x and y are the original coordinates of the joint key points, and x' and y' are the standardized coordinates of the joint key points; x pelvis and y pelvis represents the original x and y coordinates of the center point of the pelvis, shoulder_width is the Euclidean distance between the two shoulder joints, and body_height is the vertical distance from the top of the head to the ankle joint.

[0071] Finally, sliding window normalization is applied to the standardized joint coordinates across multiple frames to form a time series of standardized joint keypoint coordinates. This sliding window normalization process suppresses interference from sudden changes in motion between frames, ensuring the consistency and comparability of the output time series data.

[0072] This invention uses computer vision technology and a pose estimation algorithm to detect key points of human joints, and performs automated analysis based on a deep learning model. This avoids the problems of strong subjectivity and poor consistency caused by differences in the experience of evaluators in manual observation and evaluation, making the evaluation results more objective and accurate.

[0073] Step S2: Based on the time series of the standardized joint key point coordinates, a bidirectional recurrent neural network is used to extract the action temporal features. The target joint feature weights are enhanced in the key motion phase through a phase-aware attention mechanism. Kinematic differential modeling is performed on the motion trajectory of a specific joint, and the temporal features are encoded into a fixed-dimensional feature vector.

[0074] Specifically, a Bidirectional Long Short-Term Memory Network (Bi-LSTM) is first used to process the standardized joint keypoint coordinate time series obtained in step S1. Bi-LSTM combines forward LSTM and backward LSTM. The forward LSTM processes data from the beginning to the end of the time series, capturing positive temporal dependencies—the relationship between the current state and previous states. The backward LSTM processes data from the end to the beginning of the time series, capturing negative temporal dependencies—the relationship between the current state and subsequent states. This approach allows for a more comprehensive extraction of the temporal features of the joint keypoint coordinates.

[0075] Next, a phase-aware attention mechanism is introduced. First, kinematic features are used to identify key movement phases. Kinematic features reflect changes in various physical quantities of a joint during movement, such as position, velocity, and acceleration. By analyzing these features, significant phases of the movement can be identified, such as the takeoff phase, landing phase, and force exertion phase. The phase-aware attention mechanism automatically identifies key movement phases and enhances the feature weights of target joints based on these phases. This helps focus on the parts that have the greatest impact on motor function during complex movements, improving the specificity and effectiveness of feature extraction. For example, when evaluating running movements, the focus can be placed on the features of lower limb joints during the start and sprint phases, allowing for a more accurate analysis of joint coordination and force exertion, thus providing athletes with more targeted training recommendations.

[0076] Furthermore, attention weights for key motion phases are calculated based on the hidden states output by the Bi-LSTM bidirectional long short-term memory network. The calculation formula is as follows:

[0077] AttentionWeight = Sigmoid(W a ·h t +b a )

[0078] In the formula, W a and b a h are learnable parameters, AttentionWeight is the attention weight, and h is the attention weight. t This represents the hidden state at time step t.

[0079] The hidden state contains internal information learned by the network during time series processing, and is expressed through the learnable parameter W. a and b a And the Sigmoid function maps hidden states to attention weights. The Sigmoid function can limit the output value between 0 and 1, representing the degree of attention given to key motion phases.

[0080] Based on the calculated attention weights, the feature weights of the target joints in the critical movement phases are enhanced. The aim is to strengthen the representation of target joint features in the temporal features, because in the critical movement phases, the movement of the target joints plays a crucial role in the function and effect of the entire movement; highlighting these features helps to more accurately analyze and evaluate motor function.

[0081] Furthermore, kinematic differential modeling is performed on the motion trajectories of specific joints, and temporal features are encoded into fixed-dimensional feature vectors. This includes second-order kinematic differential modeling of the scapular motion trajectory. Second-order differential modeling can calculate the velocity and acceleration of joint motion; velocity reflects the speed of joint motion, and acceleration reflects the rate of change of joint motion velocity. These kinematic parameters can describe the joint motion state in greater depth, providing richer information for a comprehensive understanding of the movement. For example, in analyzing throwing motions, changes in scapular acceleration can reflect the process and intensity of force application. The temporal features with enhanced target joint weights are then fused with the kinematic differential modeling results of specific joints. The purpose of fusion is to integrate feature information obtained from different levels to form a more comprehensive and richer feature representation. Temporal features contain dynamic changes of the joint in the time dimension, while kinematic differential modeling results provide physical quantity information of joint motion; the two complement each other. The fused features are then input into a fully connected layer for feature compression, encoding them into fixed-dimensional feature vectors. Fully connected layers can perform linear transformations and non-linear activation operations on input features, mapping a high-dimensional feature space to a low-dimensional, fixed-dimensional space. Fixed-dimensional feature vectors facilitate subsequent processing and analysis; for example, they can be used for tasks such as classification and regression, providing standardized input for generating accurate motor function assessment results.

[0082] This invention extracts action temporal features through a bidirectional recurrent neural network, enhances target joint features in key motion phases through a phase-aware attention mechanism, and performs kinematic differential modeling of specific joint trajectories. This replaces human eye observation and single-dimensional analysis, achieving the effect of accurately capturing subtle compensatory movements, improving evaluation accuracy, and reducing missed and false judgments.

[0083] Step S3: Input the feature vector into the multi-task analysis network, extract common biomechanical features across different functional movements through the parameter sharing mechanism, perform biomechanical characteristic analysis for each functional movement, and generate quantitative evaluation results including joint range of motion, muscle compensation patterns and kinetic chain function by combining predefined biomechanical constraint rules.

[0084] Specifically, firstly, the feature vector obtained in step S2 is input into the multi-task analysis network. The purpose of the multi-task analysis network is to be able to process multiple related tasks simultaneously and improve the overall analysis efficiency and accuracy by utilizing the inherent connections between different tasks.

[0085] By leveraging a built-in parameter-sharing mechanism, common biomechanical features across different functional movements are extracted from the input feature vectors. Parameter sharing means that certain parameters in the network are the same across different tasks, avoiding the need to train a large number of parameters separately for each task, thus reducing model complexity and the risk of overfitting. Simultaneously, parameter sharing enables the network to learn common patterns and features across different functional movements, such as core stability patterns, basic joint range of motion, and kinetic chain transmission efficiency. These common features reflect the fundamental biomechanical laws of human movement and are crucial for a comprehensive understanding of human motor function.

[0086] The parameter sharing mechanism reduces the number of parameters in the model, lowers the model's complexity, avoids overfitting, and enables the model to learn common features more efficiently when dealing with different functional actions, thereby improving the overall accuracy of the analysis.

[0087] After extracting common biomechanical features, these features are reused through a parameter sharing mechanism, enabling the network to adapt and learn more quickly when handling new functional actions, thus improving the model's transferability and generalization performance. For example, when analyzing a new movement, the learned common features can be used to quickly extract key information about the action.

[0088] Furthermore, within the multi-task analysis network, specific analyses are conducted for different functional movements. For example, the overhead squat analyzes the coordination of the hip-knee-ankle joint chain, the wall angel exercise analyzes scapular trajectory stability, and the single-leg squat analyzes pelvic coronal plane control. For instance, the overhead squat focuses on analyzing the coordination of the hip-knee-ankle joint chain. Because the overhead squat involves the coordinated movement of multiple lower limb joints, the coordination of the joint chain directly affects the quality and safety of the movement. Analysis can reveal whether the movement sequence and angle changes of each joint during the movement are reasonable. For the wall angel exercise, scapular trajectory stability is analyzed. The wall angel exercise primarily trains shoulder extension and scapular control; the stability of the scapular trajectory reflects the control effect of the shoulder muscles and the coordinated movement ability of the scapula and rib cage. For the single-leg squat, pelvic coronal plane control is analyzed. During a single-leg squat, the body's center of gravity shifts, and the ability to control the pelvis in the coronal plane is crucial for maintaining balance. Analysis can assess the strength and coordination of the muscles surrounding the pelvis.

[0089] The results of specific analyses are validated using predefined biomechanical constraint rules. These rules are based on the principles of human biomechanics and extensive experimental data, such as the normal range of joint range of motion and the optimal sequence of muscle exertion. Validation determines whether the results of specific analyses conform to normal human biomechanical principles, eliminating abnormal data and erroneous analyses. For example, the formula for calculating the probability P of knee valgus is: P = σ(w1·x hip +w2·x knee The predefined constraint is w1 + w2 ≤ 1.7, ensuring that the sum of the weights of the two joint features conforms to physiological logic and avoids excessive influence of a single joint feature on the results. Here, w1 and w2 are the weights of the hip joint feature and the knee joint feature, respectively. hip x represents the characteristic value of the hip joint. knee These are characteristic values ​​of the knee joint.

[0090] Based on the validated specialized analysis results, quantitative assessment results including joint range of motion, muscle compensation patterns, and kinetic chain function are generated. Joint range of motion reflects the range of motion of a joint and is an important indicator for assessing joint function; muscle compensation patterns refer to the way other muscles or joints compensate for the movement when the function of a certain joint or muscle is limited. Understanding muscle compensation patterns helps to identify potential sports injury risks and functional abnormalities; kinetic chain function comprehensively reflects the ability of various parts of the body to work together during movement.

[0091] This invention employs a multi-task analysis network, extracting common biomechanical characteristics across different functional movements through a parameter-sharing mechanism. Simultaneously, it performs specific analyses for each movement, achieving a holistic assessment of multi-movement, multi-joint coordination patterns and comprehensively reflecting kinetic chain dysfunction. The generated quantitative assessment results include information on joint range of motion, muscle compensation patterns, and kinetic chain function, providing users with a comprehensive assessment of motor function. This information helps identify potential sports injury risks, functional abnormalities, and muscle imbalances, providing crucial evidence for developing personalized training plans and rehabilitation programs. For example, by analyzing muscle compensation patterns, it is possible to identify which muscles are overcompensating, allowing for targeted strength training and muscle balance adjustments.

[0092] Figure 2 The flowchart for assessing poor posture provided by this invention is referred to below. Figure 2 The process for assessing poor posture is as follows:

[0093] First, in the data acquisition and preprocessing stage, videos containing the subject's body posture need to be collected, and pose estimation algorithms are used to extract key joint points in each frame. Weights are adjusted based on joint detection confidence to ensure stable detection of key joints, with particular attention paid to key areas related to body posture. Next, the original coordinates of the key joint points are standardized to form time-series data.

[0094] The next step involves calculating key parameters. Firstly, angles of key body parts are calculated, such as the head tilt angle and spinal lateral tilt angle. These angles directly reflect the spatial posture of different parts of the body; for example, an abnormal head tilt angle may indicate poor posture with the head extended forward. Secondly, distance parameters, such as floor height difference and knee height difference, are calculated to help determine body symmetry and whether posture is normal.

[0095] Next, the skeletal framework information is input, which is built based on pose estimation and can more comprehensively represent the body structure. Combined with the skeletal framework, the angles and distances of key parts are calculated again to make the data more accurate.

[0096] Finally, an anomaly assessment is conducted. The calculated angles, distances, and other data are compared to a reference range for normal posture, which is similar to predefined biomechanical constraints. If the data exceeds the normal range, it is considered abnormal. A posture assessment report is generated based on the anomalies, clearly indicating the abnormal parameters and corresponding postural types, such as rounded shoulders or anterior pelvic tilt, providing a scientific basis for subsequent posture adjustments and health management.

[0097] Figure 3 The flowchart for generating a body health report provided by this invention is shown below. Figure 3 The process for generating a physical health report includes:

[0098] First, basic data is obtained according to the motor function assessment method of this invention. Through the posture estimation module, motion videos of the subject's relevant body postures are collected, and the posture estimation algorithm is used to detect key joint points in the video frames. Weights are dynamically adjusted based on the joint detection confidence level to optimize the stability of key joint detection, with particular emphasis on posture-related joints. Next, the original coordinates of the key joint points are standardized to form time-series data, providing standardized data for subsequent analysis.

[0099] Next, a risk assessment is performed, corresponding to the risk assessment section in the flowchart. The spatiotemporal feature encoding module and the multi-task decision-making module are used to analyze and assess posture. On one hand, a bidirectional recurrent neural network is used to extract temporal features of movements, and a phase-aware attention mechanism is used to enhance the features of target joints in key movement phases. Kinematic differential modeling is performed on the trajectories of specific joints to obtain posture-related feature information. On the other hand, through a multi-task analysis network and a parameter-sharing mechanism, common biomechanical features, such as core stability patterns, are extracted. Simultaneously, specific analyses are performed on movements related to different postures, and combined with predefined biomechanical constraint rules, the deviation of posture indicators is judged to determine the risk level, which is categorized as normal, mild, moderate, and severe. For example, if key angles such as spinal lateral tilt exceed the normal range, it may be judged as having a certain degree of risk.

[0100] After completing the risk assessment, the system proceeds to report generation. Based on the previously acquired postural indicators, deviations, and the assessed risk level, the system generates a detailed postural health report. The report may include specific values ​​for each postural indicator, comparisons with normal ranges, detailed deviation data, and a clear risk level. Finally, the system outputs the generated report, which users can access to clearly understand their postural health status, providing a basis for subsequent postural adjustments and health management.

[0101] Figure 4 The structural block diagram of the deep learning-based motion function assessment system provided by the present invention is shown below. Figure 4 This invention provides a deep learning-based motion function assessment system, comprising:

[0102] The pose estimation module is used to acquire motion videos of subjects performing functional movements. It uses a pose estimation algorithm to detect human joint key points in each image frame of the motion video. Based on the joint detection confidence, it dynamically adjusts the joint weights to optimize the stability of lower limb joint detection. The original coordinates of the joint key points in each frame are standardized to obtain a time series of standardized joint key point coordinates.

[0103] The spatiotemporal feature encoding module is used to extract action temporal features based on the time series of the standardized joint key point coordinates using a bidirectional recurrent neural network, enhance the target joint feature weights in key motion phases through a phase-aware attention mechanism, perform kinematic differential modeling of the motion trajectory of a specific joint, and encode the temporal features into a fixed-dimensional feature vector.

[0104] The multi-task decision module is used to input the feature vector into the multi-task analysis network, extract common biomechanical features across different functional movements through a parameter sharing mechanism, perform biomechanical characteristic analysis for each functional movement, and generate quantitative assessment results including joint range of motion, muscle compensation patterns, and kinetic chain function by combining predefined biomechanical constraint rules.

[0105] The pose estimation module is based on an improved MediaPipe BlazePose model. The core improvement of the MediaPipe BlazePose model is the adoption of a dynamic weight adjustment algorithm based on the sigmoid function. This algorithm calculates an adjusted weight in real time based on the initial confidence score of each joint point output by the model itself. First, the motion video of the subject performing functional movements is processed, and human joint keypoints are extracted from each image frame through keypoint detection. Then, the joint weights are dynamically adjusted based on the joint detection confidence score to reduce detection errors. Finally, the original coordinates of the joint keypoints in each frame are standardized, outputting a time series of standardized joint keypoint coordinates, providing accurate and unified basic data for subsequent feature encoding.

[0106] After receiving the standardized joint coordinate time series output by the pose estimation module, the spatiotemporal feature encoding module first captures the forward and backward temporal dependencies of the action through a bidirectional LSTM network to comprehensively extract the temporal correlation information of the action. Then, an attention mechanism (such as a phase-aware attention mechanism) is introduced to enhance the feature weights of the target joints during key motion phases and strengthen the expression of key motion information. Finally, the temporal feature extraction is completed, and combined with the kinematic differential modeling of the specific joint motion trajectory, the temporal features are encoded into fixed-dimensional feature vectors to achieve accurate compression and unified structured expression of action features.

[0107] After inputting a fixed-dimensional feature vector into the multi-task decision-making module, common biomechanical features across different functional movements are first extracted through a shared layer, including core stability patterns, basic joint range of motion, and kinetic chain transmission efficiency. Then, through a dedicated layer, specific biomechanical characteristic analyses are performed for each functional movement; for example, the hip-knee-ankle chain coordination is analyzed for overhead squats, and scapular trajectory stability is analyzed for wall angels. Finally, the output head, combined with predefined biomechanical constraint rules, verifies the rationality of the analysis results and generates quantitative assessment results including joint range of motion, muscle compensation patterns, and kinetic chain function, completing the multi-task collaborative assessment and result output.

[0108] The deep learning-based motor function assessment system provided by the present invention is used to execute the deep learning-based motor function assessment method provided in the foregoing embodiments. The deep learning-based motor function assessment method has been described in detail in the above embodiments, and will not be repeated here.

[0109] In a preferred embodiment, the multi-task decision module includes:

[0110] A shared layer is used to extract common biomechanical features of all functional movements, including core stability patterns, basic joint range of motion, and kinetic chain transmission efficiency.

[0111] The exclusive layer, connected to the shared layer, is used for specific analysis of various functional movements, including analyzing the coordination of the hip-knee-ankle joint chain for overhead squats, analyzing the stability of the scapular trajectory for the angel on the wall, and analyzing the coronal plane control of the pelvis for single-leg squats.

[0112] The output layer, connected to the exclusive layer, is used to integrate biomechanical constraint rules and verify the rationality of the analysis results, generating quantitative assessment results including joint range of motion, muscle compensation patterns, and kinetic chain function.

[0113] In a preferred embodiment of the present invention, a flowchart example of the deep learning-based motion function evaluation method of the present invention is provided:

[0114] Step 1: Data Acquisition and Preprocessing

[0115] 1.1 Data Acquisition Specifications

[0116] 1.1.1 Subject Preparation

[0117] Clothing requirements: Subjects should wear tight-fitting sportswear to avoid clothing obscuring joints and affecting the accuracy of key point detection.

[0118] Standing position: The subject stands in a 2.5m×2.5m calibration area with foot positioning points marked on the ground to ensure consistency of position for each data collection.

[0119] 1.1.2 Equipment Layout

[0120] Camera configuration: One 1080P RGB camera is placed in front and one on the side, with a frame rate of 30fps.

[0121] The camera was positioned at the subject's hip joint (approximately 1 meter high), with a pitch angle of 0° and a shooting distance of 2.5 meters to cover the entire range of motion.

[0122] Environmental requirements: The light intensity should be greater than 300 lux to avoid shadow interference; the background should be a solid color, and a dark green screen is recommended to improve video quality.

[0123] 1.1.3 The action execution flow is shown in Table 1.

[0124] Table 1

[0125]

[0126] Each action was repeated 3 times, and the best data was selected for subsequent analysis.

[0127] S12 Data Preprocessing Flow

[0128] S121 Video Preprocessing

[0129] The video preprocessing workflow is implemented using Python, and the pseudocode is as follows:

[0130] defpreprocess_video(video_path):

[0131] #1. Frame Extraction Processing: Extract image frames from the video at 30fps. A 15-second animation generates 450 frames, a 12-second animation generates 360 frames, and a 10-second animation generates 300 frames.

[0132] frames=extract_frames(video_path,fps=30)

[0133] #2. Background Removal: The GrabCut algorithm is used to separate the human body region from the background, eliminating background interference. frames_without_bg = grabcut_background_removal(frames)

[0134] #3. Illumination Correction: Image brightness is equalized using the CLAHE (Contrast-Limited Adaptive Histogram Equalization) algorithm to avoid uneven illumination affecting keypoint detection.

[0135] frames_corrected=clahe_light_correction(frames_without_bg)

[0136] #4. Timestamp Synchronization: Based on the frame timestamps from the dual cameras, linear interpolation is used to achieve frame alignment, ensuring that the error between frontal and lateral frames is less than 0.1 seconds.

[0137] synchronized_frames=sync_frames(front_frames=frames_corrected,side_frames=side_frames_corrected)

[0138] return synchronized_frames

[0139] 1.2.2 Key Point Detection and Verification Detection Model:

[0140] Using an improved MediaPipeBlazePose model, 17 key joints of the human body (including ankle, knee, hip, pelvis, shoulder, and top of head) are detected, and the original coordinates and detection confidence of each joint are output.

[0141] Outlier handling: 1) Confidence filtering: When detecting the confidence of a joint, cubic spline interpolation is used to repair the joint coordinates based on the joint coordinates of the preceding and following frames to avoid interference from low-confidence data. 2) Kinematic constraints: The change in knee joint angle between adjacent frames is calculated. If the angle change is greater than 30° / frame (exceeding the range of human physiological movement), the frame is re-detected to ensure the continuity of joint movement trajectory.

[0142] 1.2.3 Standardization of Key Points

[0143] Using the center point of the pelvis as the origin of the local coordinate system, the differences in shooting angle and individual standing position are eliminated; using the Euclidean distance between the two shoulder joints as the horizontal dimension benchmark and the vertical distance from the top of the head to the ankle joint as the vertical dimension benchmark, the differences in individual body shape are eliminated.

[0144] 1.2.4 Sliding Window Normalization

[0145] To eliminate individual gait rhythm differences, a sliding window normalization process is applied to the continuously standardized joint coordinates: the window size is set to 15 frames, corresponding to a 0.5-second motion segment, covering the smallest motion unit. The dynamic adjustment formula is as follows: Here, Δ is the normalized value, which makes the numerical distribution of the same action phases of different subjects comparable.

[0146] 1.3 Quality Control Mechanism

[0147] 1.3.1 Data Validation

[0148] Spatial verification: Based on the laws of human kinematics, the deviation of the angle of the line connecting the hip, knee and ankle joints is verified (the normal range is 170°-180°). If the deviation is greater than 10°, the data of that frame is deemed invalid and discarded.

[0149] Timing verification: Perform FFT (Fast Fourier Transform) analysis on the action frame sequence to extract the main frequency of the action. If the main frequency is not in the range of 0.5-2Hz (the normal frequency range of human functional actions), the action is judged to be non-standard and needs to be re-acquired.

[0150] 1.3.2 Sample Augmentation Strategies

[0151] To improve the robustness of the model, sample augmentation is performed on the preprocessed data. The specific parameters are shown in Table 2 below:

[0152] Table 2

[0153]

[0154] 1.3.3 Data Storage Specifications

[0155] A hierarchical directory structure is used to store the collected and preprocessed data to ensure data traceability.

[0156] 1.4 Technical Advantages and Verification

[0157] 1.4.1 Comparison of Pretreatment Efficiency

[0158] The data preprocessing process of this invention is compared with the data processing method of traditional motor function assessment. The results are shown in Table 3 below. This invention is significantly superior to the traditional method in both processing speed and accuracy.

[0159] Table 3

[0160]

[0161] 1.4.2 Clinical validation results

[0162] Reliability test: Three senior physical therapists (with ≥5 years of experience) independently labeled the data of 120 subjects, and the intragroup correlation coefficient (ICC) was calculated with the output of this system. The results showed that ICC = 0.91 (p<0.001), indicating that the evaluation results of this system are highly consistent with the professional human evaluation.

[0163] Validity test: Using the Vicon optical motion capture system (the industry gold standard) as a reference, the joint angle detection results of this system were compared with those of Vicon. The results showed that the joint angle error was 1.2°±0.3°, which is less than the clinically acceptable threshold of 1.5°, thus verifying the assessment validity of this system.

[0164] 2. Model Training Details

[0165] The core analytical model of this invention is a "multi-task decision model". The training specifications, technical parameters and performance verification of this model are described in detail below to ensure that the model can be trained stably and meet the needs of clinical evaluation.

[0166] 2.1 Standardized naming for training models

[0167] To avoid ambiguity in the documentation, the core model used for motion function assessment in this invention is uniformly named "multi-task decision model" in all scenarios. This model covers loss function design, multi-task decision modules (shared layer, exclusive layer) and constraint output layer, and is responsible for converting the feature vectors output by the spatiotemporal feature encoding module into quantitative assessment results.

[0168] 2.2 Detailed Explanation of Training Architecture

[0169] 2.2.1 Model Composition Structure

[0170] The multi-task decision model consists of three parts: a shared layer, an exclusive layer, and a constraint output layer. The functions of each component and their corresponding system modules are shown in Table 4 below.

[0171] Table 4

[0172]

[0173]

[0174] 2.2.2 Complete Model Flow

[0175] The input and output flow of the multi-task decision model is as follows, realizing end-to-end processing from raw data to evaluation results:

[0176] Posture estimation module → Standardized joint coordinates → Spatiotemporal feature encoding module (outputs 256-dimensional feature vector) → (Shared layer → Exclusive layer → Constrained output layer) → Quantitative evaluation results (joint range of motion, muscle compensation pattern, kinetic chain function).

[0177] 2.3 Training Technique Details

[0178] 2.3.1 Loss Function Design

[0179] A multi-task weighted cross-entropy loss function is used to balance the training weights of the three types of actions, as shown in the following formula:

[0180] L = 0.4L overhead +0.3L wall +0.3L singleleg

[0181] L task=-∑y ture log(p pred )

[0182] In the formula, L overhead L wall and L singleleg The cross-entropy losses for overhead squats, wall angels, and single-leg squats are respectively, y ture For action tags, such as knee valgus; p pred Predict probabilities for the model.

[0183] Weighting criteria: Through cross-validation with 120 subjects, the overhead squat, which involves a multi-joint chain of hip-knee-ankle, has the highest motion complexity and is therefore assigned the highest weight (0.4); the wall angel and the single-leg squat have similar complexity and are both assigned a weight of 0.3 to ensure that the model evaluates each movement with balanced accuracy.

[0184] 2.3.2 Optimizer Configuration

[0185] The NAdam optimizer (Adam optimizer with Nesterov momentum) is used to balance convergence speed and stability.

[0186] 2.3.3 Key Training Strategies

[0187] To improve the training efficiency and generalization ability of the multi-task decision-making model, the following training strategy is adopted, and the specific parameters are shown in Table 5 below:

[0188] Table 5

[0189]

[0190] 2.4 Training Data Stream and Hardware Deployment

[0191] 2.4.1 Training Data Stream

[0192] A pipelined data processing workflow is adopted to ensure efficient and stable training. Figure 5 As shown.

[0193] 2.4.2 Hardware Configuration Requirements

[0194] To meet the requirements of model training and real-time inference, the hardware configuration is divided into minimum and recommended configurations. The specific parameters and training time are shown in Table 6 below:

[0195] Table 6

[0196]

[0197] 2.4.3 Performance Verification Indicators

[0198] After training, the performance of the multi-task decision model was tested, and the results are shown in Table 7 below. All indicators meet the requirements for clinical application.

[0199] Table 7

[0200]

[0201]

[0202] 2.5 Integration with System Architecture

[0203] The multi-task decision model is the core decision-making unit of the motion function assessment system of this invention, and its relationship with other modules of the system is as follows:

[0204] Input dependency: Receive a 256-dimensional fixed-dimensional feature vector output by the spatiotemporal feature encoding module. This vector is obtained by fusing and compressing the temporal features extracted by bidirectional LSTM with the results of joint kinematic differential modeling.

[0205] Output correlation: Generate quantitative assessment reports for three types of movements. The reports include: ① Joint range of motion score (e.g., hip flexion range of motion 90°, score 8 / 10); ② Probability of muscle compensation pattern (e.g., knee valgus compensation probability 85%); ③ Kinematic chain dysfunction markers (e.g., low hip-knee-ankle chain transmission efficiency).

[0206] Real-time performance guarantee: On the recommended hardware configuration (RTX3090), the model can achieve real-time evaluation at 30fps (matching the camera frame rate), meeting the needs of "action execution - immediate evaluation" in clinical scenarios.

[0207] 2.6 Clinical validation results

[0208] A double-blind clinical trial was conducted on 120 subjects (including 60 healthy individuals and 60 individuals undergoing sports injury rehabilitation). The clinical applicability of the system was verified by comparing the evaluation results of the present invention with those of experienced therapists (≥5 years of experience). The results are shown in Table 8 below:

[0209] Table 8

[0210]

[0211]

[0212] The results show that the system of the present invention is superior to traditional manual assessment in terms of assessment accuracy, efficiency and consistency of results, and can meet the actual needs of clinical motor function assessment.

[0213] Figure 6 A structural block diagram of the electronic device provided by the present invention, such as Figure 6As shown, the present invention also provides an electronic device 600, which can be a mobile terminal, desktop computer, laptop, handheld computer, server, or other computing device. The electronic device 600 includes a processor 601 and a memory 602, wherein the memory 602 stores a deep learning-based motion function evaluation program 603.

[0214] In some embodiments, memory 602 may be an internal storage unit of a computer device, such as a hard disk or memory. In other embodiments, memory 602 may be an external storage device of a computer device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, memory 602 may include both internal and external storage units of the computer device. Memory 602 is used to store application software and various types of data installed on the computer device, such as program code for installing the computer device. Memory 602 can also be used to temporarily store data that has been output or will be output. In one embodiment, when the deep learning-based motion function evaluation program 603 is executed by processor 601, the following steps are implemented:

[0215] The study acquires motion videos of subjects performing functional movements, uses a pose estimation algorithm to detect human joint key points in each image frame of the motion video, dynamically adjusts joint weights based on joint detection confidence to optimize the stability of lower limb joint detection, and standardizes the original coordinates of joint key points in each frame to obtain a time series of standardized joint key point coordinates.

[0216] Based on the time series of the standardized joint key point coordinates, a bidirectional recurrent neural network is used to extract the action temporal features. The target joint feature weights are enhanced in the key motion phase through a phase-aware attention mechanism. Kinematic differential modeling is performed on the motion trajectory of a specific joint, and the temporal features are encoded into a fixed-dimensional feature vector.

[0217] The feature vectors are input into a multi-task analysis network, and common biomechanical features across different functional movements are extracted through a parameter sharing mechanism. At the same time, biomechanical characteristic analysis is performed on each functional movement, and quantitative evaluation results including joint range of motion, muscle compensation patterns, and kinetic chain function are generated by combining predefined biomechanical constraint rules.

[0218] In some embodiments, processor 601 may be a central processing unit (CPU), microprocessor or other data processing chip, used to run program code stored in memory 602 or process data, such as executing a deep learning-based motion function evaluation program.

[0219] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

[0220] 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 of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based method for assessing motor function, characterized in that, include: We acquire motion videos of subjects performing functional movements, use pose estimation algorithms to detect key points of human joints in each image frame of the motion video, and dynamically adjust joint weights based on joint detection confidence to optimize the stability of lower limb joint detection. The original coordinates of the joint key points in each frame are standardized to obtain a time series of standardized joint key point coordinates. Based on the time series of the standardized joint key point coordinates, a bidirectional recurrent neural network is used to extract the action temporal features. The target joint feature weights are enhanced in the key motion phase through a phase-aware attention mechanism. Kinematic differential modeling is performed on the motion trajectory of a specific joint, and the temporal features are encoded into a fixed-dimensional feature vector. The feature vectors are input into a multi-task analysis network, and common biomechanical features across different functional movements are extracted through a parameter sharing mechanism. At the same time, biomechanical characteristic analysis is performed on each functional movement, and quantitative evaluation results including joint range of motion, muscle compensation patterns, and kinetic chain function are generated by combining predefined biomechanical constraint rules.

2. The deep learning-based motor function assessment method according to claim 1, characterized in that, The method employs a pose estimation algorithm to detect key human joints in each image frame of a motion video, and dynamically adjusts joint weights based on joint detection confidence to optimize the stability of lower limb joint detection. Specifically, this includes: Obtain the initial confidence level C for each joint. j The weights of each joint are dynamically calculated using the Sigmoid function; the weight calculation formula is as follows: In the formula, W j C is the adjusted weight of joint j. j Let θ be the initial confidence level of joint j, k be the slope, and θ be the confidence threshold. Based on the aforementioned weights, priority is given to improving the detection stability of lower limb joints, ensuring the reliability of the detection results of key points of lower limb joints.

3. The deep learning-based motor function assessment method according to claim 1, characterized in that, The process of standardizing the original coordinates of joint key points in each frame to obtain a time series of standardized joint key point coordinates specifically includes: Using the center point of the pelvis as the origin of the local coordinate system, obtain the original coordinates of the center point of the pelvis; The original coordinates of joint keypoints in each frame are standardized using the Euclidean distance between the two shoulder joints as the lateral dimension reference and the vertical distance from the top of the head to the ankle joint as the longitudinal dimension reference. The standardization formula is as follows: In the formula, x and y are the original coordinates of the joint key points, and x' and y' are the standardized coordinates of the joint key points; x pelvis and y pelvis represents the original x and y coordinates of the center point of the pelvis, shoulder_width is the Euclidean distance between the two shoulder joints, and body_height is the vertical distance from the top of the head to the ankle joint. A sliding window normalization process is performed on the standardized joint coordinates of multiple consecutive frames to form a time series of standardized joint key point coordinates.

4. The deep learning-based motor function assessment method according to claim 1, characterized in that, The step of extracting motion temporal features using a bidirectional recurrent neural network based on the time series of the standardized joint key point coordinates, and enhancing the target joint feature weights during key motion phases through a phase-aware attention mechanism, specifically includes: A bidirectional long short-term memory network is used to process the time series of standardized joint key point coordinates. Forward LSTM captures the forward temporal dependency, and backward LSTM captures the backward temporal dependency. A phase-aware attention mechanism is introduced to identify key motion phases of an action through kinematic features; the attention weights of the key motion phases are calculated based on the hidden states output by the bidirectional long short-term memory network, using the following formula: AttentionWeight=Sigmoid(W a ·h t +b a ) In the formula, W a and b a h are learnable parameters, AttentionWeight is the attention weight, and h is the attention weight. t The hidden state at time step t; The attention weights enhance the feature weights of target joints in key motion phases, thereby strengthening the expression of target joint features in temporal features.

5. The deep learning-based motor function assessment method according to claim 1, characterized in that, The process of performing kinematic differential modeling on the motion trajectory of a specific joint and encoding temporal features into a fixed-dimensional feature vector specifically includes: Second-order kinematic differential modeling is performed on the scapular motion trajectory to calculate joint motion velocity and acceleration; The temporal features enhanced with target joint weights are fused with the kinematic differential modeling results of specific joints, and then input into a fully connected layer for feature compression, encoding them into fixed-dimensional feature vectors.

6. The deep learning-based motor function assessment method according to claim 1, characterized in that, The step of inputting the feature vector into a multi-task analysis network and extracting common biomechanical features across different functional actions through a parameter sharing mechanism specifically includes: The feature vector is input into a multi-task analysis network. Through the network's built-in parameter sharing mechanism, common biomechanical features across different functional actions are extracted, and these common biomechanical features are reused through the parameter sharing mechanism. The common biomechanical characteristics include core stability mode, basic joint range of motion, and kinetic chain transmission efficiency. Core stability mode is a characteristic that reflects the coordinated control ability of the trunk and pelvis. Basic joint range of motion is a characteristic that reflects the basic range of motion of the hip and shoulder joints. Kinetic chain transmission efficiency is a characteristic that reflects the effectiveness of force transmission from the lower limbs to the upper limbs.

7. The deep learning-based motor function assessment method according to claim 1, characterized in that, The biomechanical characteristic analysis of each functional movement, combined with predefined biomechanical constraint rules, generates quantitative assessment results including joint range of motion, muscle compensation patterns, and kinetic chain function, specifically including: In the multi-task analysis network, specific analyses are performed for different functional movements. For the overhead squat, the coordination of the hip-knee-ankle joint chain is analyzed; for the angel on the wall, the stability of the scapular trajectory is analyzed; and for the single-leg squat, the coronal plane control of the pelvis is analyzed. The rationality of the specific analysis results is verified by using predefined biomechanical constraint rules. Based on the verified specific analysis results, quantitative assessment results including joint range of motion, muscle compensation patterns and kinetic chain function are generated.

8. A deep learning-based motor function assessment system, the system being used to execute the deep learning-based motor function assessment method according to any one of claims 1-7, characterized in that, include: The pose estimation module is used to acquire motion videos of subjects performing functional movements. It uses a pose estimation algorithm to detect key points of human joints in each image frame of the motion video and dynamically adjusts joint weights based on joint detection confidence to optimize the stability of lower limb joint detection. The original coordinates of the joint key points in each frame are standardized to obtain a time series of standardized joint key point coordinates. The spatiotemporal feature encoding module is used to extract action temporal features based on the time series of the standardized joint key point coordinates using a bidirectional recurrent neural network, enhance the target joint feature weights in key motion phases through a phase-aware attention mechanism, perform kinematic differential modeling of the motion trajectory of a specific joint, and encode the temporal features into a fixed-dimensional feature vector. The multi-task decision module is used to input the feature vector into the multi-task analysis network, extract common biomechanical features across different functional movements through a parameter sharing mechanism, perform biomechanical characteristic analysis for each functional movement, and generate quantitative assessment results including joint range of motion, muscle compensation patterns, and kinetic chain function by combining predefined biomechanical constraint rules.

9. The deep learning-based motion function assessment system according to claim 8, characterized in that, The multi-task decision module includes: A shared layer is used to extract common biomechanical features of all functional movements, including core stability patterns, basic joint range of motion, and kinetic chain transmission efficiency. The exclusive layer, connected to the shared layer, is used for specific analysis of various functional movements, including analyzing the coordination of the hip-knee-ankle joint chain for overhead squats, analyzing the stability of the scapular trajectory for the angel on the wall, and analyzing the coronal plane control of the pelvis for single-leg squats. The output layer, connected to the exclusive layer, is used to integrate biomechanical constraint rules and verify the rationality of the analysis results, generating quantitative assessment results including joint range of motion, muscle compensation patterns, and kinetic chain function.

10. An electronic device, Its features are, Including memory and processor, among which, The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the deep learning-based motion function evaluation method according to any one of claims 1 to 7.

Images