An adaptive attitude capture and biomechanical feature analysis method and system

Abstract

This invention provides an adaptive posture capture and biomechanical feature analysis method and system, comprising the following steps: constructing a posture detection network based on an improved YOLOv8-Pose architecture; replacing traditional static convolutions with full-dimensional dynamic convolutions ODConv in the backbone and neck networks; dynamically weighting spatial, channel, and convolution kernel weights using a multi-dimensional attention mechanism to achieve real-time detection of athlete postures and capture of skeletal key points, outputting preliminary raw coordinates of skeletal key points; introducing a geometric calibration logic based on biomechanical constraints, utilizing the invariance of bone length in short-term motion sequences to correct the preliminary detected skeletal key points with bone length consistency constraints and geometric constraints, obtaining calibrated key points, and outputting kinematic evaluation parameters based on the calibrated key points. This invention significantly improves posture estimation accuracy while maintaining real-time feedback capabilities, providing a scientific kinematic evaluation basis for sports research and practical training.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of sports engineering and computer vision, specifically relating to an adaptive posture capture and biomechanical feature analysis method and system. Background Technology

[0002] With the rapid development of computer vision and deep learning technologies, video-based markerless human pose estimation methods have been gradually applied to the analysis of athletes' technical movements and training feedback. Current mainstream pose estimation algorithms, such as OpenPose, HRNet, and the YOLO-Pose series, mostly employ neural network structures with static convolutional kernels, exhibiting good detection accuracy in typical scenarios.

[0003] However, in the field of sports science, when athletes perform high-level competitive movements (such as the 100-meter sprint, high jump, and racket-like sports), their bodies are in a state of rapid transition between extreme eccentric and concentric contractions, accompanied by severe non-rigid deformation and motion blur in the limbs. The fixed receptive field and weight distribution of static convolutional kernels actually assume the stationarity of target features, which means that when processing high-angular-velocity joint movements, the network cannot dynamically adjust its feature extraction strategy according to changes in the target's shape, resulting in the "keypoint drift" phenomenon.

[0004] Taking fast-swinging sports like table tennis, tennis, and badminton as examples, athletes' joint angular velocities can reach hundreds of degrees per second when performing actions such as forehand attacks, serves, or smashes, resulting in significant non-rigid deformation of the limbs. Simultaneously, motion blur and limb occlusion (such as the arm obscuring the torso during a swing) are common problems during image acquisition. Traditional static convolutional networks exhibit particularly noticeable keypoint detection drift when processing such high-speed dynamic targets, especially at distal nodes such as the wrist and ankle.

[0005] Such minute shifts in key points, which might be overlooked in routine motion recognition, can lead to errors of several degrees in joint angle calculations during high-level sports training. For example, in badminton smash analysis, a slight deviation in the elbow joint angle can affect the judgment of power efficiency; in table tennis topspin, an error in the knee joint angle can lead to a misunderstanding of the weight transfer mechanism.

[0006] Therefore, a high-precision attitude capture method that can adapt to motion state and take into account biomechanical rationality is needed to solve the key point drift and feature extraction ambiguity problems in the current technology when dealing with high-speed, non-rigid moving targets, and to meet the needs of sports scientific research and actual training for precise motion analysis. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides an adaptive posture capture and biomechanical feature parsing method and system. The invention aims to construct a YOLOv8-Pose posture detection network improved based on full-dimensional dynamic convolution ODConv, and utilize a multi-dimensional attention mechanism to adaptively calibrate motion features, effectively suppressing keypoint drift. Simultaneously, it introduces a biomechanical geometric calibration strategy based on bone length consistency constraints to post-process and optimize the initially detected keypoints, eliminating anatomical anomalies and ensuring that the output skeletal data conforms to the biomechanical principles of human movement.

[0008] This invention significantly improves the accuracy of posture estimation while maintaining real-time feedback capability, especially enhancing the detection stability of key points at the extremities, providing a scientific and refined kinematic assessment basis for sports research and practical training.

[0009] The present invention achieves the above-mentioned technical objectives through the following technical means.

[0010] An adaptive attitude capture and biomechanical feature parsing method includes the following steps:

[0011] Step S1: Construct a pose detection network based on YOLOv8-Pose improvement. In the backbone network and neck network, full-dimensional dynamic convolution ODConv is used to replace the traditional static convolution. Multi-dimensional attention mechanism is used to dynamically weight the space, channel and convolution kernel weights to realize the real-time detection of the athlete's movement posture and capture of skeletal key points, and output the initial original coordinates of the skeletal key points.

[0012] Step S2: Introduce a geometric calibration logic based on biomechanical constraints. Utilize the invariance of bone length in short-time motion sequences to perform bone length consistency constraints and geometric constraint corrections on the bone key points initially detected in Step S1, obtain calibrated key points, and output kinematic evaluation parameters based on the calibrated key points.

[0013] In the above scheme, step 1 specifically includes the following steps:

[0014] Step S1.1: Use high-definition camera equipment or load standard datasets to acquire images of athletes, preprocess the acquired images, and construct a standardized training dataset;

[0015] Step S1.2: Input the image data into the improved YOLOv8-Pose network for training, and perform supplementary incremental training for samples with missed detections or key point drift.

[0016] Step S1.3: Input the preprocessed image data into the improved YOLOv8-Pose algorithm network structure for feature extraction and multi-scale feature fusion to identify the key points of the human skeleton. The improved YOLOv8-Pose algorithm network structure integrates full-dimensional dynamic convolution ODConv.

[0017] Step S1.4: Using the improved YOLOv8-Pose algorithm network structure, obtain the bounding box coordinates and the original coordinate set of skeletal key points for each athlete.

[0018] Furthermore, the preprocessing described in step S1.1 includes uniform size adjustment, normalization, and Mosaic data augmentation. Furthermore, the improved YOLOv8-Pose algorithm network structure described in step S1.3 includes a backbone network...

[0019] The network consists of a neck and a head. During the downsampling phase of the backbone and at key connections in the neck, full-dimensional dynamic convolution ODConv is used instead of traditional static convolution Conv.

[0020] In the above scheme, the key points of the human skeleton mentioned in step S1.3 are 17.

[0021] In the above scheme, the ODConv module utilizes a multi-dimensional attention mechanism to perform parallel deep learning on the input feature map across four dimensions: spatial scale, input channels, output channels, and convolutional kernel weights; its dynamic weight aggregation formula is as follows:

[0022] in, For the generated dynamic convolution kernel, The attention scalar representing space. The attention scalar representing the input channel. The attention scalar representing the output channel. The attention scalar represents the weights of the convolutional kernel. For the first A static convolutional kernel, This represents multiplication operations along different dimensions.

[0023] Furthermore, the total loss function during model training Includes classification loss, bounding box loss, and pose key point loss. The loss of key points The OKS Loss is calculated using the following formula:

[0024]

[0025] in, To predict the Euclidean distance between keypoints and ground truth points, The target scale factor, For the first Normalization constants at key points, As a visibility marker for key points, This is an indicator function used to determine whether a key point is included in the calculation.

[0026] In the above scheme, step S2 includes the following steps:

[0027] Step S2.1: Construct a human skeleton vector model and define the length of the bones. The Euclidean distance between adjacent keypoints; Step S2.2: Implement a skeleton length consistency constraint, utilizing the fact that the length of the human skeleton remains constant in a short-term motion sequence.

[0028] Based on biological characteristics, a reference length threshold is set, and when the bone length of a certain frame is detected... Mutations exceeding a set threshold When the keypoint is determined to have drifted, geometric constraint correction is performed using the correct position from the previous frame to obtain the calibrated keypoint. ;

[0029] Step S2.3: Calculate kinematic parameters based on the calibrated key points and output a comprehensive training evaluation report including joint angles, motion scores and keyframe images.

[0030] Furthermore, the joint angle mentioned in step S2.3 includes the knee joint angle. The calculation formula is as follows:

[0031]

[0032] in, These represent the coordinates of the hip, knee, and ankle joints, respectively. and This is the corresponding joint vector.

[0033] A system for implementing the adaptive posture capture and biomechanical feature analysis method includes an image acquisition module, a posture detection module, a post-processing optimization module, and a kinematic analysis module.

[0034] The image acquisition module is used to acquire motion images of athletes;

[0035] The pose detection module incorporates a pose detection network based on YOLOv8-Pose. This improved pose detection network uses full-dimensional dynamic convolution ODConv instead of traditional static convolution in the backbone and neck networks. It utilizes a multi-dimensional attention mechanism to dynamically weight the spatial, channel, and convolution kernel weights to extract features and detect key points in the input image, and outputs the initial coordinates of the skeletal key points.

[0036] The post-processing optimization module is used to perform geometric calibration on the preliminary key points based on biomechanical constraints, and to use the invariance of bone length in short-time motion sequences to perform bone length consistency constraints and geometric constraint corrections, and output the calibrated bone key points.

[0037] The kinematic analysis module is used to calculate kinematic parameters based on calibrated skeletal key points and output a comprehensive training evaluation report.

[0038] Compared with the prior art, the beneficial effects of the present invention are:

[0039] This invention addresses the issues of keypoint drift and blurred feature extraction in existing pose estimation algorithms when dealing with high-speed, non-rigid moving targets. First, it constructs a feature extraction network based on an improved YOLOv8-Pose architecture. A full-dimensional dynamic convolutional ODConv module is introduced into the key nodes of the backbone network and feature pyramid. A multi-dimensional attention mechanism is used to dynamically weight the spatial, channel, and convolutional kernel weights, adaptively calibrating the motion features. Second, a geometric calibration strategy based on human biomechanical constraints is introduced. Utilizing the invariance of bone length in short-term motion sequences, secondary geometric constraint corrections are applied to the initially detected keypoints, eliminating anatomical anomalies.

[0040] This invention effectively suppresses keypoint drift, improves the average accuracy of attitude estimation to 45.2%, and increases inference latency by only 0.1ms, maintaining real-time feedback capability. At the same time, it can accurately calculate kinematic parameters such as joint angles and angular velocities, providing a scientific evaluation basis for sports research and practical training. The table below shows the relevant specific implementation data.

[0041] This invention addresses motion fuzzing in high-speed, non-rigid moving targets (such as racket-swinging actions in table tennis, tennis, and badminton).

[0042] In complex scenarios such as limb occlusion, this invention can accurately capture key points of the human skeleton and effectively suppress key point drift. The improved YOLOv8-Pose algorithm model of this invention performs excellently on the MPII public dataset. This dataset covers 25 sports movement categories, including running, jumping, throwing, and ball sports, exhibiting high scene diversity. Experimental results show that this invention has good detection capabilities for different types of sports movements, laying the foundation for subsequent applications in specific sports scenarios (such as table tennis, tennis, and badminton). The model can adapt to complex conditions such as different lighting, shooting angles, and sports types, possessing strong transfer application potential. It is suitable for posture capture and analysis tasks of various high-speed, non-rigid motion targets, and can perform posture capture and biomechanical feature analysis in various sports scenarios, providing coaches with scientific training evaluation basis. Attached Figure Description

[0043] Figure 1 This is the overall operational logic diagram of the adaptive attitude capture and biomechanical feature analysis method of the present invention;

[0044] Figure 2 This is a diagram of the improved YOLOv8 algorithm network structure of this invention;

[0045] Figure 3 This is a diagram of the ODConv network module of the present invention. Detailed Implementation

[0046] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0047] The present invention will be further described in detail below with reference to the embodiments. This embodiment uses the MPII dataset and sports-specific data as the verification basis, and the experimental environment is shown in Table 1. The input image data uniformly adopts 640 The model was standardized to 640 pixels, with a model size of m. The optimizer was selected as the stochastic gradient descent (SGD) algorithm, with an initial learning rate of 0.01 and an cyclic learning rate of 0.01. The model was trained for a total of 450 epochs.

[0048] Table 1 Experimental Environment Setup

[0049]

[0050] like Figure 1As shown, an adaptive posture capture and biomechanical feature parsing method is a deep learning-based adaptive posture capture and biomechanical feature parsing method for high-speed, non-rigid motion targets. The aim is to propose a high-precision athlete posture estimation network while enhancing robustness to motion blur and limb occlusion. Through specialized training and optimization for high-speed motion scenarios, the model effectively improves the recognition accuracy of key points at the human extremities (such as the wrist and ankle) in fast-swinging movements in sports like table tennis, tennis, and badminton. This improves the accuracy of athlete motion capture, increases motion analysis efficiency, and optimizes the training feedback process. The method includes the following steps:

[0051] Step S1: Based on the YOLOv8-Pose improved algorithm with a full-dimensional dynamic perception mechanism, real-time detection and key point capture of the athlete's movement posture are achieved.

[0052] A pose detection network based on the improved YOLOv8-Pose is constructed. Full-dimensional dynamic convolution ODConv is used in the backbone and neck network to replace traditional static convolution. A multi-dimensional attention mechanism is used to dynamically weight the spatial, channel, and kernel weights to achieve real-time detection of athlete's movement posture and capture of skeletal key points, outputting the initial raw coordinates of the skeletal key points. The specific steps include:

[0053] Step S1.1: Image Acquisition and Dataset Construction

[0054] Images of athletes are captured using high-definition cameras or by loading the MPII standard dataset. The captured images are preprocessed to construct a standardized training dataset. To ensure the model is highly robust to sports scenarios, the captured data must cover different lighting conditions (indoor and outdoor), different shooting angles (front, side, and top views), and different types of sports (running, jumping, and throwing). The captured images are preprocessed, including uniform size adjustment, normalization, and Mosaic data augmentation, to construct a standardized training dataset.

[0055] Step S1.2, Model Training and Incremental Training

[0056] After completing the dataset construction, the image data was input into the improved YOLOv8-Pose network for training. During the training process, the weights were tested in multiple rounds to evaluate the model's performance under conditions of large limb deformation (such as the moment of take-off in a high jump) and severe occlusion (such as group defense). Supplementary incremental training was performed on samples with missed detections or key point drift to optimize the model's sensitivity to specific sports movements. Incremental training refers to supplementing the training on the basis of the already trained model for specific difficult samples, rather than retraining from scratch.

[0057] Step 1.2.1: After completing the dataset construction, input the image data into the improved YOLOv8-Pose network for initial training. Set the training epochs to 100 and the learning rate to [value missing]. Set the batch size to 16.

[0058] Step 1.2.2: After initial training, use the current model to perform inference on the validation set and automatically filter difficult samples according to the following criteria: Missed samples: Confidence of human detection boxes. Keypoint drift samples: OKS values ​​for pose estimation Severely occluded samples: Number of visible key points 17 key points (more than 7 out of 17 key points are invisible); 285 difficult samples were selected from 2710 validation set images based on the above criteria. The samples with significant deformation were defined as having a bone length variance exceeding 20% ​​of the reference length.

[0059] Step 1.2.3: Perform targeted data augmentation on the selected difficult samples to generate more similar training samples. Apply Random Erasing to randomly occlude 10%~30% of the image region; apply Affine Transform with a rotation angle of ±15° and a scaling ratio of 0.8~1.2; apply Motion Blur with a kernel size of 5×5; generate 3~5 augmented copies for each difficult sample, expanding the number of difficult samples to approximately 1000.

[0060] Step 1.2.4: Construct an incremental training dataset containing 70% original training data + 30% hard samples and their augmented copies. Adjust training hyperparameters: Learning rate: Reduce to (1 / 10 of the initial learning rate) to avoid destroying learned features; Batch size: reduced to 8 to enhance the model's memory of difficult samples; Number of training epochs: set to 20 epochs for short-term fine-tuning; Loss weight adjustment: increase the loss weight of pose keypoints from 12.0 to 15.0 to enhance the optimization of keypoint localization; Network freezing strategy: freeze the parameters of the first 50 layers of the network and only fine-tune the high-level semantic feature layers.

[0061] Step 1.2.5: Repeat steps 1.2.2 to 1.2.4 to perform multiple rounds of iterative incremental training until one of the following convergence conditions is met: the improvement of mAP@0.5 is less than 0.5%; the average OKS value of hard samples reaches more than 0.75; and 3 rounds of incremental training are completed.

[0062] Step S1.3: Feature Extraction and Key Point Recognition

[0063] The preprocessed image data is input into the improved YOLOv8-Pose algorithm network structure for feature extraction and multi-scale feature fusion, and 17 skeletal key points of the human body are identified. The improved YOLOv8-Pose algorithm network structure integrates full-dimensional dynamic convolution ODConv.

[0064] The improved YOLOv8-Pose algorithm network structure includes a backbone network, a neck network, and a head network, such as... Figure 2 As shown, in the downsampling stage of the backbone (layers P3, P4, P5) and at the critical connections of the neck, full-dimensional dynamic convolution ODConv is used instead of the traditional static convolution Conv. As a specific network configuration example, this invention replaces the 3rd, 5th, and 7th convolutional layers with full-dimensional dynamic convolution ODConv in the backbone network; and replaces the 16th and 19th convolutional layers with full-dimensional dynamic convolution ODConv in the neck network. This aims to introduce a dynamic mechanism at the bottleneck of information compression, achieving optimal results with minimal cost. Furthermore, an OKS loss function specifically for pose estimation is introduced during training.

[0065] The input layer standardizes the image before inputting it into the feature extraction network. The backbone network uses a CSP structure combined with the ODConv module, such as... Figure 3 As shown, the ODConv module utilizes a multi-dimensional attention mechanism to perform parallel deep learning on the input feature map across four dimensions: spatial scale, input channels, output channels, and convolutional kernel weights. Its dynamic weight aggregation formula is as follows:

[0066] in, For the generated dynamic convolution kernel, The attention scalar representing space. The attention scalar representing the input channel. The attention scalar representing the output channel. The attention scalar represents the weights of the convolutional kernel. For the first A static convolutional kernel, This represents multiplication operations along different dimensions.

[0067] The full-dimensional dynamic convolution (ODConv) module generates a four-dimensional attention scalar through the following steps:

[0068] Global context extraction: extracting the input feature map Perform global average pooling to compress the spatial dimension to To obtain the global context vector .

[0069] Multi-branch attention computation: Using the global context vector Four independent fully connected network branches are input respectively. After dimensionality reduction, ReLU activation, dimensionality increase, and Sigmoid normalization, four attention scalars are output respectively: Attention to the number of convolutional kernels, for control Weight allocation for each candidate convolutional kernel; Output channel attention, which controls the importance of each output feature channel; Input channel attention controls the importance of each input feature channel. Spatial attention controls the weights of each spatial location within the convolution kernel.

[0070] Dynamic convolution kernel generation: combining four attention scalars with The pre-learned static convolutional kernels are weighted and aggregated dimension by dimension to generate a dynamic convolutional kernel that adapts to the current input. .

[0071] The above mechanism enables the network to dynamically adjust its feature extraction strategy based on the characteristics of the input motion image (such as the amplitude of the motion and the complexity of the background), thereby improving its ability to capture complex motion postures.

[0072] The four attention scalars are calculated differently, as follows:

[0073] Convolution kernel weights and attention Calculate independently for each convolution kernel:

[0074] The purpose of this attention scalar is to... candidate convolutional kernels Weights are assigned to determine which convolutional kernel(s) should be used preferentially for the current input. Therefore, The output is a Dimensional vector:

[0075]

[0076] in, Indicates the first The weight coefficients of each convolutional kernel satisfy the following condition: (Achieved through Softmax or normalization).

[0077] The other three attention scalars are shared globally:

[0078] Output channel attention Input channel attention Spatial attention These kernels are shared across all candidate convolutional kernels. Their function is to globally modulate the internal structure of the convolutional kernels.

[0079] Modulate the input channel weights of all convolutional kernels to suppress channels containing background noise;

[0080] Modulate the output channel weights of all convolutional kernels to enhance the feature channels related to the target;

[0081] Modulate the spatial location weights of all convolution kernels to adapt the receptive field to the direction of limb extension.

[0082] The Neck network employs a PANet architecture to fuse feature maps of different scales and also incorporates the ODConv module to enhance the fusion of deep semantic features (such as action type) and shallow positional features (such as joint coordinates). The Head detection head is responsible for ultimately regressing heatmaps of 17 keypoints. The total loss function during model training is... Includes classification loss, bounding box loss, and pose key point loss. ;

[0083] The key point loss The OKS Loss is calculated using the following formula:

[0084]

[0085] in, To predict the Euclidean distance between keypoints and ground truth points, The target scale factor, For the first Normalization constants at key points, As a visibility marker for key points, This is an indicator function used to determine whether a keypoint is included in the calculation. Target scale factor. Calculated based on the bounding box size of the athlete in the image:

[0086]

[0087] in, and These represent the width and height of the athlete's bounding box (in pixels). This factor is used to normalize the detection error between athletes captured at different distances, ensuring fair evaluation of large near-field targets and small far-field targets at the same scale. Keypoint normalization constant. The value of is a predefined constant based on the statistical characteristics of the dataset annotations, reflecting the first... The inherent uncertainty of key points during manual annotation. This invention employs... The values ​​are based on the standard settings of the COCO / MPII dataset, and the specific values ​​are shown in Table 2 below:

[0088] Table 2. Standard settings for the COCO / MPII dataset.

[0089]

[0090] This invention uses a 17-keypoint human skeleton model in COCO format. The keypoint index is defined as shown in Table 3 below:

[0091] Table 3 Definition of Key Points Index of Human Skeleton

[0092]

[0093] Total loss function during model training Includes classification loss, bounding box loss, and pose key point loss. :

[0094]

[0095] in, This weight configuration follows the official recommendations of YOLOv8-Pose, where... This configuration, accounting for 60% of the total weight, reflects the design philosophy of this invention, which focuses on precise localization of key points as the core optimization objective. Experiments on the MPII dataset have yielded excellent results.

[0096] Step S1.4: Output the original keypoint coordinates

[0097] The bounding box coordinates for each athlete are obtained using the improved YOLOv8-Pose algorithm network structure. and the original coordinate set of skeletal key points .

[0098] Step S2: Geometric calibration and kinematic parameter calculation based on biomechanical constraints

[0099] A geometric calibration logic based on biomechanical constraints is introduced to perform post-processing optimization on the key points detected in the initial step and output kinematic evaluation parameters: by utilizing the invariance of bone length in short-time motion sequences, the bone key points detected in the initial step S1 are corrected by bone length consistency constraints and geometric constraints to obtain calibrated key points, and kinematic evaluation parameters are output based on the calibrated key points.

[0100] Step S2 specifically includes the following steps:

[0101] Step S2.1: Construct a human skeleton vector model

[0102] Based on the topological structure of human anatomy, the first... Root bone length Adjacent key points and The Euclidean distance between them is given by the following formula:

[0103]

[0104] Step S2.2, Bone Length Consistency Constraint

[0105] Implement bone length consistency constraints, leveraging the biological characteristic that human bone length remains constant in short-term motion sequences, and set a reference length threshold. When the bone length of a certain frame is detected... Mutations exceeding a set threshold Time (i.e.) If the keypoint is determined to have drifted, the following coordinate correction rules are executed. For ease of description, let's call it... Let this be the reference length of the bone segment in the absence of drift. These are the keypoint coordinates corresponding to the nearest valid frame that has not experienced drift in time.

[0106] Rule 1: Use the keypoint coordinates corresponding to the nearest valid frame that has not experienced drift in time sequence. As a correction benchmark.

[0107] Rule 2: Drift key points of the current frame Position constraints are applied along the direction of the line connecting it to the reference point, and the corrected coordinates are obtained. It is located on this line.

[0108] Rule 3: Revised Key Points With reference point The Euclidean distance does not exceed the reference length of the bone segment. ,Right now If the current detection point The distance to the reference point is less than or equal to Then keep No stretching correction is performed.

[0109] If drift occurs across multiple consecutive frames, the most recent valid frame is recursively used as the reference. This ultimately yields the calibrated set of skeletal keypoints. .

[0110] Step S2.3: Calculation and evaluation of kinematic parameters and output of the evaluation report.

[0111] Based on calibrated keypoints, kinematic parameters are calculated, and a comprehensive training evaluation report including joint angles, movement scores, and keyframe images is output to guide coaches in conducting scientific training. For example, calculating knee joint angles... To assess the standardization of movement. Select the hip joint. Knee joint Ankle The coordinates are used to construct a vector. and The angle is calculated using the dot product formula:

[0112]

[0113] in, These represent the coordinates of the hip, knee, and ankle joints, respectively. and This is the corresponding joint vector.

[0114] The key point mapping for mainstream joint angle calculation is shown in Table 4.

[0115] Table 4. Joints with computable angles and their keypoint mappings

[0116]

[0117] Example verification:

[0118] The joint angle calculation verification is shown in Table 5. The knee joint angle was calculated based on the test results and compared with the standard range. The joint angle reference standards for different types of sports are shown in Table 6, which lists the standard angle ranges for movements such as sprinting, squatting, and throwing.

[0119] Table 5 Examples of knee joint angle calculation for sprint starting action

[0120]

[0121] Table 6. Joint Angle Reference Standards for Typical Movements

[0122]

[0123] Movement scoring and level assessment: Scores are given based on joint angles and movement standardization, and the assessment levels are shown in Table 7.

[0124] Table 7. Action Scoring Levels and System Feedback

[0125]

[0126] Performance verification:

[0127] The performance comparison, model complexity, and inference speed of this invention on the MPII dataset are shown in Tables 8 and 9.

[0128] Table 8 compares the performance of the model before and after the improvement on the MPII dataset:

[0129] This invention effectively suppresses keypoint drift, improving the average accuracy of attitude estimation from 43.6% to 45.2%, while increasing inference latency by only 0.1ms, maintaining real-time feedback capability; at the same time, it can accurately calculate kinematic parameters such as joint angles and angular velocities, providing a scientific evaluation basis for sports research and practical training.

[0130] The model complexity and real-time performance are verified as shown in Table 9. The inference latency only increased by 0.1ms, maintaining real-time performance.

[0131] Table 8. Performance comparison of the model before and after improvement on the MPII dataset.

[0132]

[0133] Table 9 Comparison of Model Complexity and Inference Speed

[0134]

[0135] The above data are all based on real experimental results on the MPII dataset, which fully demonstrates the beneficial effects of this invention in terms of accuracy improvement, real-time performance maintenance, and kinematic parameter calculation. The MPII Human Pose Dataset is an internationally recognized standard test set for human pose estimation, released by the Max Planck Institute in Germany. It contains approximately 25,000 images covering 410 everyday and sports activity scenarios. This dataset already includes samples of ball sports such as tennis and football; therefore, the validation of this invention on MPII indirectly covers some specific sports scenarios.

[0136] A system for implementing the adaptive posture capture and biomechanical feature analysis method includes an image acquisition module, a posture detection module, a post-processing optimization module, and a kinematic analysis module.

[0137] The image acquisition module is used to acquire motion images of athletes;

[0138] The pose detection module incorporates a pose detection network based on YOLOv8-Pose. This improved pose detection network uses full-dimensional dynamic convolution ODConv instead of traditional static convolution in the backbone and neck networks. It utilizes a multi-dimensional attention mechanism to dynamically weight the spatial, channel, and convolution kernel weights to extract features and detect key points in the input image, and outputs the initial coordinates of the skeletal key points.

[0139] The post-processing optimization module is used to perform geometric calibration on the preliminary key points based on biomechanical constraints, and to use the invariance of bone length in short-time motion sequences to perform bone length consistency constraints and geometric constraint corrections, and output the calibrated bone key points.

[0140] The kinematic analysis module is used to calculate kinematic parameters based on calibrated skeletal key points and output a comprehensive training evaluation report.

[0141] This invention employs full-dimensional dynamic convolution (ODConv) instead of static convolution in the key downsampling nodes of the YOLOv8-Pose backbone network and feature pyramid. Through multi-dimensional dynamic weighting of space, channels, and the number of convolutional kernels, the network can adaptively adjust its receptive field and feature response, effectively suppressing keypoint drift under high-speed motion, especially improving the detection accuracy of end nodes such as the wrist and ankle. Simultaneously, biomechanical constraints based on bone length consistency are introduced in the post-processing stage to perform geometric calibration on the initial keypoints and remove anatomical anomalies. Experiments on the MPII dataset show that the average pose estimation accuracy (mAP50) of this invention is improved from 43.6% to 45.2%, with an inference latency increase of only 0.1 ms (from 1.3 ms / image to 1.4 ms / image), maintaining real-time feedback capability and providing accurate kinematic parameter evaluation for sports training.

[0142] It should be understood that although this specification is described according to various embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0143] The detailed descriptions listed above are merely specific illustrations of feasible embodiments of the present invention and are not intended to limit the scope of protection of the present invention. All equivalent embodiments or modifications made without departing from the spirit of the present invention should be included within the scope of protection of the present invention.

Claims

1. An adaptive attitude capture and biomechanical feature analysis method, characterized in that, Includes the following steps: Step S1: Construct a pose detection network based on YOLOv8-Pose improvement. In the backbone network and neck network, full-dimensional dynamic convolution ODConv is used to replace the traditional static convolution. Multi-dimensional attention mechanism is used to dynamically weight the space, channel and convolution kernel weights to realize the real-time detection of the athlete's movement posture and capture of skeletal key points, and output the initial original coordinates of the skeletal key points. Step S2: Introduce a geometric calibration logic based on biomechanical constraints. Utilize the invariance of bone length in short-time motion sequences to perform bone length consistency constraints and geometric constraint corrections on the bone key points initially detected in Step S1, obtain calibrated key points, and output kinematic evaluation parameters based on the calibrated key points.

2. The adaptive attitude capture and biomechanical feature analysis method according to claim 1, characterized in that, Step 1 specifically includes the following steps: Step S1.1: Use high-definition camera equipment or load standard datasets to acquire images of athletes, preprocess the acquired images, and construct a standardized training dataset; Step S1.2: Input the image data into the improved YOLOv8-Pose network for training, and perform supplementary incremental training for samples with missed detections or key point drift. Step S1.3: Input the preprocessed image data into the improved YOLOv8-Pose algorithm network structure for feature extraction and multi-scale feature fusion to identify the key points of the human skeleton. The improved YOLOv8-Pose algorithm network structure integrates full-dimensional dynamic convolution ODConv. Step S1.4: Using the improved YOLOv8-Pose algorithm network structure, obtain the bounding box coordinates and the original coordinate set of skeletal key points for each athlete.

3. The adaptive attitude capture and biomechanical feature analysis method according to claim 2, characterized in that, The preprocessing described in step S1.1 includes uniform size adjustment, normalization, and Mosaic data augmentation.

4. The adaptive attitude capture and biomechanical feature analysis method according to claim 2, characterized in that, The improved YOLOv8-Pose algorithm network structure described in step S1.3 includes a backbone network, a neck network, and a head detection network. In the downsampling stage of the backbone and at the key connections of the neck, the traditional static convolution Conv is replaced by full-dimensional dynamic convolution ODConv.

5. The adaptive attitude capture and biomechanical feature analysis method according to claim 2, characterized in that, The key points of the human skeleton mentioned in step S1.3 are 17.

6. The adaptive attitude capture and biomechanical feature analysis method according to claim 4, characterized in that, The ODConv module utilizes a multidimensional attention mechanism to perform parallel deep learning on the input feature map in four dimensions: spatial scale, input channel, output channel, and convolutional kernel weights. The dynamic weight aggregation formula is as follows: in, For the generated dynamic convolution kernel, The attention scalar representing space. The attention scalar representing the input channel. The attention scalar representing the output channel. The attention scalar represents the weights of the convolution kernel. For the first A static convolutional kernel, This represents multiplication operations along different dimensions.

7. The adaptive attitude capture and biomechanical feature analysis method according to claim 4, characterized in that, Total loss function during model training Includes classification loss, bounding box loss, and pose key point loss. The key point loss The OKS Loss is calculated using the following formula: in, To predict the Euclidean distance between keypoints and ground truth points, The target scale factor, For the first Normalization constants at key points, As a visibility marker for key points, This is an indicator function used to determine whether a key point is included in the calculation.

8. The adaptive attitude capture and biomechanical feature analysis method according to claim 1, characterized in that, Step S2 includes the following steps: Step S2.1: Construct a human skeleton vector model and define the length of the bones. The Euclidean distance between adjacent keypoints; Step S2.2: Implement bone length consistency constraints. Utilizing the biological characteristic that human bone length remains constant in short-term motion sequences, a reference length threshold is set. When a bone length in a certain frame is detected... Mutations exceeding a set threshold When the keypoint is determined to have drifted, geometric constraint correction is performed using the correct position from the previous frame to obtain the calibrated keypoint. ; Step S2.3: Calculate kinematic parameters based on the calibrated key points and output a comprehensive training evaluation report including joint angles, motion scores and keyframe images.

9. The adaptive attitude capture and biomechanical feature analysis method according to claim 8, characterized in that, The joint angles mentioned in step S2.3 include the knee joint angle. The calculation formula is as follows: in, These represent the coordinates of the hip, knee, and ankle joints, respectively. and This is the corresponding joint vector.

10. A system for implementing the adaptive attitude capture and biomechanical feature analysis method according to any one of claims 1-9, characterized in that, It includes an image acquisition module, a pose detection module, a post-processing optimization module, and a kinematics analysis module; The image acquisition module is used to acquire motion images of athletes; The pose detection module incorporates a pose detection network based on YOLOv8-Pose. This improved pose detection network uses full-dimensional dynamic convolution ODConv instead of traditional static convolution in the backbone and neck networks. It utilizes a multi-dimensional attention mechanism to dynamically weight the spatial, channel, and convolution kernel weights to extract features and detect key points in the input image, and outputs the initial coordinates of the skeletal key points. The post-processing optimization module is used to perform geometric calibration on the preliminary key points based on biomechanical constraints, and to use the invariance of bone length in short-time motion sequences to perform bone length consistency constraints and geometric constraint corrections, and output the calibrated bone key points. The kinematic analysis module is used to calculate kinematic parameters based on calibrated skeletal key points and output a comprehensive training evaluation report.

Images