A sports training based motion capture method

Abstract

The application belongs to the technical field of sports training, and provides a motion capture method based on sports training, which comprises the following steps: setting an integrated data acquisition network comprising a collection unit group and an intelligent node, performing athlete identity matching and binding on the collection unit group, collecting image data and athlete sensing basic data sets in real time within a motion capture period, evaluating the confidence state of the collection unit in the collection unit group, if the confidence state is reliable, calculating a high-precision three-dimensional posture based on the athlete sensing basic data sets, otherwise, using an inverse kinematics algorithm to complete the posture, integrating to generate IMU calculation information, obtaining the image data to judge the occlusion state of the athlete in real time, marking the occlusion period, dynamically selecting and executing the corresponding data processing strategy according to whether the athlete is in the occlusion period, combining the image data and the IMU calculation information to generate the optimal posture estimation of the athlete, and arranging to obtain a complete posture action sequence.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of sports training, and in particular relates to a motion capture method based on sports training. BACKGROUND

[0002] In the field of sports training, motion capture technology, as an important tool in the field of sports science, provides data support for training optimization through accurate recording of athlete motion trajectories and quantitative analysis of kinematic parameters. In the fields of competitive sports and public fitness, accurate motion analysis can help athletes optimize motion patterns, improve athletic performance, and effectively reduce the risk of sports injuries. With the continuous improvement of sports competition level, the demand for scientific training is increasing, and the application range of motion capture technology in the sports field is expanding. Traditional motion capture schemes are mainly divided into two categories: machine vision-based and inertial measurement unit (IMU)-based.

[0003] Machine vision-based motion capture can provide high spatial positioning accuracy. However, this type of motion capture has strict requirements for the environment, and its capture effect is easily affected by severe occlusion. Once the athlete himself or others, equipment causes occlusion, data loss will occur. In the multi-person collaborative training scene, this problem is particularly prominent. IMU-based motion capture systems have the advantages of being wireless and not being limited by environmental light and space, and have good application prospects in outdoor and complex sites. However, IMU systems have inherent sensor drift problems. The pose information obtained by integral operation will accumulate errors over time, resulting in a decrease in long-term accuracy, making it difficult to meet the needs of long-term, high-precision training analysis. In addition, in the multi-person training scene, how to accurately bind the data collected by multiple sensor units to the identity of a specific athlete, i.e., the data tracing problem, is also a technical difficulty. The existing simple binding method is easy to cause data confusion due to slight displacement of the sensor or signal crossing during intense exercise.

[0004] To solve the above problems, the application provides a motion capture method based on sports training. SUMMARY

[0005] In order to make up for the deficiencies of the prior art and solve at least one technical problem proposed in the background art.

[0006] The technical scheme adopted by the application to solve the technical problem is: a motion capture method based on sports training, comprising:

[0007] An integrated data acquisition network including a set of acquisition units and intelligent nodes is set up, the set of acquisition units is matched and bound to the identity of an athlete, and image data and a set of athlete sensor basic data are collected in real time during a motion capture period;

[0008] The confidence state of the acquisition unit in the acquisition unit group is evaluated, if the confidence state is reliable, the high-precision three-dimensional posture is calculated based on the sensor basis data set of the athlete, otherwise the inverse kinematics algorithm is used for posture completion, and the IMU calculation information is integrated to generate;

[0009] The image data is acquired to determine the shielding state of the athlete in real time, and the shielding period is marked;

[0010] According to whether it is in the shielding period, the corresponding data processing strategy is dynamically selected and executed, the optimal posture estimation of the athlete is generated by combining the image data and the IMU calculation information, and the complete posture action sequence is obtained;

[0011] Further, the specific way of athlete identity matching binding is:

[0012] An identity feature library containing the apparent feature vector is constructed for each athlete;

[0013] The integrated data acquisition network acquires inertial data, coordinate data and image data, the coordinate data includes the three-dimensional position coordinates of the acquisition unit in the acquisition unit group in the global coordinate system, the image data includes the node field of view image collected by the high-definition camera of each intelligent node, the candidate search area is generated on the corresponding node field of view image based on the coordinate data of the acquisition unit group, the target detection algorithm is run in the candidate search area to locate the pixel-level bounding box of the athlete, the apparent feature vector is extracted by constructing and using the feature extraction model, and the identity feature library is matched, the cosine similarity is calculated as a matching quantitative index, and the optimal allocation algorithm is applied. The principle of the highest matching quantitative index is used to realize the athlete identity matching binding;

[0014] Further, the generation method of the candidate search area is:

[0015] The coordinate data of the acquisition unit group is acquired, the coordinate data is projected to the two-dimensional image plane of the node field of view image collected by the corresponding high-definition camera through the camera extrinsic parameter matrix of the used high-definition camera, the corresponding pixel coordinate point is obtained, and the candidate search area of fixed size is generated in the node field of view image in combination with the preset athlete body shape ratio parameter;

[0016] Further, the evaluation method of the confidence state is:

[0017] Based on the pre-defined human body skeleton model and biomechanical constraints, reasonable physiological ranges are set for the relative distance and angle between the acquisition units in each limb segment, the confidence state of the acquisition unit is initialized and marked as reliable, and the relative position relationship between each acquisition unit is continuously monitored. When the relative position relationship continuously exceeds the reasonable physiological range within the upper limit of the preset deviation length, the confidence state of the corresponding acquisition unit is marked as unreliable;

[0018] Further, the manner of solving the high-precision three-dimensional pose is:

[0019] The sensor-based data set of the athlete includes inertial data and coordinate data collected by the corresponding set of acquisition units, the inertial data including angular velocity, acceleration, and magnetic field vector, a sliding time window is set, the inertial data in the window is sorted into an inertial time sequence and preprocessed, a sensor fusion algorithm based on complementary filtering is used to fuse and solve the inertial data, and a quaternion update equation is used to output the high-precision three-dimensional pose of the acquisition unit in the global coordinate system;

[0020] Further, the manner of completing the pose is:

[0021] A predefined human skeleton model and the high-precision three-dimensional poses of all acquisition units with a confidence state of being reliable in the set of acquisition units are used, an inverse kinematics solver based on an optimization algorithm is used to minimize the error between the actual high-precision three-dimensional pose of the acquisition unit and the calculated pose of the corresponding joint of the human skeleton model as an objective function, and a numerical optimization method is used to iteratively calculate the optimal pose of all joints of the human skeleton model in real time, and the high-precision three-dimensional pose of the acquisition unit with a confidence state of being unreliable is deduced and predicted;

[0022] Further, the manner of marking the occlusion period is:

[0023] The image data includes a node field of view image collected by a high-definition camera of each intelligent node;

[0024] The coordinate data of the athlete is projected into the node field of view image to generate a candidate search area, an occlusion evaluation model based on deep learning is used to identify the candidate search area, and an occlusion type is output, the occlusion type including an incomplete athlete contour, an occlusion overlap, or no occlusion, if all candidate search areas generated for the athlete are identified as no occlusion, the athlete is determined to be in a non-occluded state, otherwise, the athlete is determined to be occluded, and a period in which the athlete is occluded is marked as an occlusion period;

[0025] Further, the data processing strategy includes a tightly coupled fusion strategy, and the specific manner is:

[0026] Based on the node field of view image of the athlete in the image data, the three-dimensional coordinates of the key joints of the athlete are calculated by binocular vision triangulation as visual solving information, a Kalman filtering algorithm is used to accurately calibrate the long-term error in the IMU solving information caused by sensor drift accumulation, and the optimal pose estimation of the athlete at the current time is output;

[0027] Further, the data processing strategy also includes an estimation backtracking correction strategy, and the specific manner is:

[0028] The preliminary pose estimation of the athlete in the occlusion period is estimated, and at the end of the occlusion period, the preliminary pose estimation of the athlete in the occlusion period is arranged in time sequence into an occlusion pose estimation sequence, an optimization problem is constructed by combining the occlusion pose estimation sequence and the visual solution information at the end of the occlusion period, the optimization problem takes minimizing the visual re-projection error and the pose smoothness as a joint objective function, and a batch optimization adjustment is performed through a nonlinear least squares optimization algorithm to obtain the optimal pose estimation of the athlete in the occlusion period.

[0029] Further, the estimation mode of the preliminary pose estimation is:

[0030] The coordinate data in the athlete's sensor basic data set is taken as a space constraint, and the IMU solution information of the athlete is optimized and solved under a unified inverse kinematics framework to estimate the preliminary pose estimation of the athlete in the occlusion period.

[0031] The beneficial effects of the present application are as follows:

[0032] 1. The present application significantly improves the accuracy and robustness of motion capture by constructing an integrated data acquisition network and realizing tight coupling and deep fusion of multi-source data, and places the visual information of the global perspective and the inertial measurement data of the local details under a unified space-time reference for collaborative processing. In the unoccluded period, the long-term drift error of the IMU sensor caused by integral operation can be directly and effectively calibrated using high-precision joint coordinates based on binocular vision triangulation solution, which fundamentally overcomes the inherent defects of pure inertial navigation system. This deep fusion mechanism ensures that the system can still output stable, smooth and cumulative error-free high-precision three-dimensional pose even under intense motion during the entire motion capture period, providing a very reliable data analysis basis for coaches.

[0033] 2. The present application has intelligent fault tolerance and adaptive processing capability, ensuring the continuity and integrity of data acquisition in complex real training environment. By real-time evaluation of the confidence state of the acquisition unit and the visual occlusion state of the athlete, abnormal conditions such as sensor displacement, failure or visual occlusion can be intelligently identified. On this basis, the system dynamically triggers the corresponding inverse kinematics pose completion or estimation backtracking correction strategy, so that during the sensor abnormality or temporary complete occlusion period, physically reasonable and coherent pose estimation can still be generated. This active fault tolerance mechanism effectively avoids the common data interruption or jump problem in traditional methods, greatly improving the practical value and reliability of the system in real training. BRIEF DESCRIPTION OF DRAWINGS

[0034] The present application will be further described below with reference to the accompanying drawings.

[0035] Figure 1is a step flow chart of a motion capture method based on sports training according to an embodiment of the present application;

[0036] Figure 2 is a specific acquisition step flow chart of a complete posture motion sequence in a motion capture method based on sports training according to an embodiment of the present application. DETAILED DESCRIPTION

[0037] In order to make the technical means, creative features, purposes and effects of the present application easy to understand, the present application will be further described below in conjunction with specific embodiments.

[0038] Embodiment 1

[0039] Please refer to Figure 1 The motion capture method based on sports training according to an embodiment of the present application aims to solve the key problems of data confusion in multi-person training scenarios, inherent drift of inertial sensors, data interruption caused by displacement faults, and data loss caused by visual occlusion in the prior art. By constructing an integrated data acquisition network and performing athlete identity matching and binding to ensure data traceability, by evaluating the confidence state of the acquisition unit and fusing high-precision posture solving and inverse kinematics algorithm to ensure the reliability and completeness of the data, by real-time judging the occlusion state and dynamically selecting the tight coupling fusion strategy or the estimation backtracking correction strategy to realize the intelligent fusion and complement of multi-source information, the complete posture motion sequence of the athletes is generated in the complex real training environment, which is continuous, high-precision and high-robustness. Specifically, the following steps are included:

[0040] S1: Set up an integrated data acquisition network including an acquisition unit group and an intelligent node, perform athlete identity matching and binding on the acquisition unit group, and real-time collect image data and athlete's sensor basic data group in the motion capture period;

[0041] An integrated data acquisition network for sports training motion capture is set up, which is composed of an acquisition unit group equipped on each athlete's body and an intelligent node deployed in the training field. The acquisition unit group includes multiple acquisition units, and the acquisition units equipped on the same athlete form an acquisition unit group.

[0042] Specifically, the equipped sites of the acquisition units are planned according to human kinematics at key joints of the human body, and the joint identifiers corresponding to the joints are assigned. Lightweight acquisition units are provided for athletes, each acquisition unit integrates a group of IMU (Inertial Measurement Unit) sensor groups and an active UWB (Ultra-Wideband) tag. Multiple intelligent nodes are deployed around the training field and in key areas, each intelligent node integrates a high-definition camera and a UWB base station module.

[0043] The IMU sensor group includes a three-axis accelerometer, a three-axis gyroscope and a three-axis magnetometer, and the collection unit is directly fixed to a planned equipped site on the athlete's limb under the premise of ensuring comfort, and is used to collect inertial data reflecting local limb movement, including angular velocity, acceleration and magnetic field vector.

[0044] The UWB tag and the UWB base station jointly constitute a UWB positioning system, the time difference of UWB signals arriving at different UWB base stations is measured, the three-dimensional position coordinates of the collection unit in the global coordinate system are solved, and the coordinate data of each collection unit group is obtained.

[0045] The high-definition camera is used to collect image data reflecting the global movement of the athlete, and the image data includes the node field of view image collected by the high-definition camera of each intelligent node, and the collection of the image data covers the entire training field, and each position in the training field is covered by the node field of view image of two high-definition cameras.

[0046] It should be noted that the UWB positioning system not only provides spatial position information, but also provides a time reference for the entire system through a precise clock signal.

[0047] The period of motion capture of the athlete in the training field is marked as a motion capture period, and at the start of the motion capture period, the athlete identity matching and binding of the collection unit group is performed.

[0048] Specifically, an exclusive identity feature library is constructed for each athlete, and specifically, a unique identity code is assigned to each athlete, and the high-definition camera is controlled to collect multiple frames of standard posture images of the front, side and specific angle of each athlete under the condition of no occlusion and good illumination, and a feature extraction model based on a deep convolutional neural network is constructed and used to extract an apparent feature vector of the body shape contour and morphological features of the effective coded athlete from the collected multiple frames of standard posture images, and the extracted apparent feature vector is bound with the corresponding unique identity code and stored in the identity feature library.

[0049] At the start of the motion capture period, for any collection unit group, the coordinate data of the collection unit group is obtained, the coordinate data is projected to the two-dimensional image plane of the node field of view image collected by the corresponding high-definition camera through the pre-calibrated camera extrinsic matrix of the high-definition camera, the corresponding pixel coordinate point is obtained, and a candidate search region of a fixed size is generated in the node field of view image in combination with the pre-set athlete body size ratio parameter.

[0050] running a target detection algorithm in the candidate search area, locating the pixel-level bounding box of the matched athlete of the acquisition unit group, extracting the appearance feature vector in the pixel-level bounding box by using a feature extraction model, and using a feature similarity calculation algorithm to compare the extracted appearance feature vector with all appearance feature vectors in the identity feature library one by one, calculating the cosine similarity as a matching quantitative indicator, and applying an optimal assignment algorithm to match a unique identity code for the athlete in the pixel-level bounding box based on the principle that the matching quantitative indicator is the highest, so as to realize the identity matching and binding of the athletes of the acquisition unit group;

[0051] For any athlete, the inertial data and coordinate data collected by the athlete corresponding acquisition unit group are arranged into a sensor basic data group of the athlete, and during the motion capture period, the integrated data acquisition network performs data acquisition based on a unified space-time reference and a preset acquisition frequency, to obtain real-time image data and the sensor basic data group of each athlete;

[0052] It should be noted that the purpose of this step is to synchronously acquire image data and sensor basic data groups through the integrated data acquisition network, and to accurately associate the data collected by the acquisition unit group with a specific athlete in a multi-person training scene through athlete identity matching and binding, thereby avoiding data confusion and ensuring the accuracy of data traceability;

[0053] S2: If the confidence state of the acquisition unit in the acquisition unit group is trusted, a high-precision three-dimensional pose is calculated based on the sensor basic data group of the athlete, otherwise, a reverse kinematics algorithm is used for pose completion, and IMU calculation information is integrated and generated;

[0054] For any athlete, the sensor basic data group of the athlete is acquired in real time, and the relative position relationship between each acquisition unit is calculated, a pre-defined human body skeletal model and biomechanical constraints are used to set a reasonable physiological range for the relative distance and angle between acquisition units in each limb segment, the confidence state of the acquisition unit is initialized and marked as trusted, and the relative position relationship between each acquisition unit is continuously monitored. When the relative position relationship continuously exceeds the reasonable physiological range within the upper limit of the preset deviation duration, it is determined that the corresponding acquisition unit has displacement or internal failure, and the confidence state of the acquisition unit is marked as untrusted;

[0055] It should be noted that the confidence state of the acquisition unit will not be automatically restored to the trusted state after being marked as untrusted, and the confidence state needs to be initialized manually after the position of the acquisition unit is manually restored;

[0056] For the acquisition unit with a trusted confidence state, the inertial data collected by the acquisition unit in the sensor basic data group of the athlete is processed for high-precision pose calculation, and a high-precision three-dimensional pose represented in the form of a quaternion is output;

[0057] Specifically, a sliding time window with a current time as an end point is set to slide over time, inertia data in the sliding time window is arranged in time sequence as an inertia time sequence, and the inertia time sequence is preprocessed including outlier rejection, moving average filtering and zero offset compensation to eliminate noise and errors in the inertia time sequence;

[0058] For the preprocessed inertia time sequence, a sensor fusion algorithm based on complementary filtering is used to fuse and solve the inertia data, including angular velocity, acceleration and magnetic field vector. The angular velocity in the inertia time sequence is time integrated to generate a dynamic attitude estimate. The acceleration and magnetic field vector are processed by vector normalization and orthogonalization to respectively solve the pitch angle, roll angle and yaw angle of the collection unit relative to the gravity direction and the geomagnetic north direction to construct an absolute attitude reference matrix. The dynamic attitude estimate and the absolute attitude reference matrix are fused by the complementary filtering algorithm, and the fusion contribution is adjusted by the pre-set filter coefficient weighting. The high-precision three-dimensional attitude of the collection unit in the global coordinate system is output by the quaternion update equation, and the high-precision three-dimensional attitude is represented in the form of quaternion;

[0059] It should be noted that the high-precision three-dimensional attitude represents the spatial orientation state of the limb segment in the global coordinate system accurately described by the inertia data of the collection unit with a confidence state of being trusted;

[0060] For the collection unit with a confidence state of being untrusted, a pre-defined human body skeletal model and the high-precision three-dimensional attitude of all collection units with a confidence state of being trusted in the collection unit group are used to complete the attitude by inverse kinematics algorithm;

[0061] Specifically, the pre-defined human body skeletal model is a tree-like topological structure including joint nodes, limb segments and their degree of freedom constraints, which defines the biomechanical parameters such as the length of each limb segment and the joint range of motion. The high-precision three-dimensional attitude of all collection units with a confidence state of being trusted in the collection unit group is obtained, and the joint node identifier of the corresponding collection unit in the human body skeletal model is obtained as the input of the inverse kinematics algorithm. An inverse kinematics solver based on an optimization algorithm is used, the inverse kinematics solver takes minimizing the error between the actual high-precision three-dimensional attitude of the collection unit and the calculated attitude of the corresponding joint node of the human body skeletal model as the objective function, and through numerical optimization method, the optimal pose of all joint nodes in the human body skeletal model is calculated in real time, and the high-precision three-dimensional attitude of the collection unit with a confidence state of being untrusted is predicted and derived;

[0062] All high-precision three-dimensional attitudes of the athletes are integrated and marked as IMU solution information;

[0063] It should be noted that the role of this step is to automatically identify sensor displacement or failure by evaluating the confidence state of the acquisition unit, and trigger the corresponding fault-tolerant processing mechanism to ensure the reliability of data acquisition. For the acquisition unit with an untrusted confidence state, inverse kinematics algorithm is used for posture completion to ensure the completeness and continuity of IMU solution information, avoiding the problem of data loss of the entire limb chain due to sensor displacement or failure.

[0064] S3: Obtain image data to determine the occlusion state of the athlete in real time, and mark the occlusion period;

[0065] An occlusion state is set for each athlete and initialized as no occlusion. The occlusion state is updated in real time based on the node view image in the image data. For any athlete, the coordinate data is projected onto the two-dimensional image plane of the node view image collected by the high-definition camera to obtain the corresponding pixel coordinate point. A candidate search region of a fixed size is generated in the node view image by combining the preset athlete body size ratio parameter;

[0066] An occlusion evaluation model based on deep learning is constructed and used to evaluate the occlusion state of the athlete. The input is the candidate search region, and the output is the identified occlusion type, including incomplete athlete outline, occlusion overlap, or no occlusion. The occlusion evaluation model is trained using a large number of athlete images with the same size as the candidate search region and corresponding occlusion type labels as samples;

[0067] If all the candidate search regions generated for the athlete are identified as no occlusion, the athlete's occlusion state is determined as no occlusion. Otherwise, the athlete's occlusion state is determined as being occluded. The athlete's occlusion state is updated in real time, and the period when the athlete's occlusion state is being occluded is marked as the occlusion period.

[0068] It should be noted that this step is to determine the occlusion state of the athlete in real time and automatically based on image data, directly identifying the occlusion type rather than simply judging target loss, providing more fine-grained occlusion information.

[0069] S4: Dynamically select and execute the corresponding data processing strategy according to whether it is in the occlusion period, generate the optimal pose estimation of the athlete by combining the image data and IMU solution information, and arrange to obtain the complete pose action sequence;

[0070] As shown in Figure 2 The specific acquisition steps of the complete pose action sequence are as follows.

[0071] Specifically, if not in the occlusion period, the data processing strategy is a tight coupling fusion strategy, adopts a tight coupling fusion algorithm, based on the image data of the corresponding node field of view of the athlete, the three-dimensional coordinates of the key joint of the athlete are calculated by binocular vision triangulation as visual calculation information, the visual calculation information and the IMU calculation information are deeply fused, the Kalman filter algorithm is adopted, the long-term error of the IMU calculation information caused by the sensor drift accumulation is accurately calibrated by the visual calculation information, and the optimal attitude estimation of the athlete at the current time is output;

[0072] For the occlusion period, the data processing strategy is an estimation backtracking correction strategy;

[0073] Specifically, in the occlusion period, the coordinate data of the athlete is used as the spatial constraint of the overall attitude of the athlete, and the IMU calculation information of the athlete is optimized and solved in a unified inverse kinematics framework to estimate the preliminary attitude estimation of the athlete in the occlusion period;

[0074] At the end of the occlusion period, the backtracking correction process is started, the preliminary attitude estimation of the athlete in the occlusion period is arranged in time sequence as an occlusion attitude estimation sequence, and the occlusion attitude estimation sequence and the visual calculation information at the end of the occlusion period are combined to construct an optimization problem, the optimization problem takes the minimization of the visual re-projection error and the attitude smoothness as the joint objective function, and the nonlinear least squares optimization algorithm is used to optimize and adjust the occlusion attitude estimation sequence in batches to obtain the optimal attitude estimation of the athlete in the occlusion period;

[0075] At the end of the motion capture period, the optimal attitude estimation of the athlete at each time in the motion capture period is arranged in time sequence to obtain a complete attitude motion sequence generated by motion capture of the athlete's sports training;

[0076] It should be noted that the purpose of this step is to use the data processing strategy dynamically selected according to the occlusion period, to pursue the highest accuracy by using the tight coupling fusion strategy in the absence of occlusion, and to ensure data continuity by using the estimation backtracking correction strategy in the presence of occlusion, to realize motion capture with high robustness throughout the process, in the tight coupling fusion strategy, the long-term error of the IMU calculation information is accurately calibrated by using the visual calculation information, which fundamentally solves the drift problem of pure inertial measurement, and the estimation backtracking correction strategy uses the reliable visual calculation information after the end of the occlusion period to correct the entire occlusion attitude estimation sequence through batch optimization, which improves the accuracy of the data during the occlusion period;

[0077] The technical scheme of the embodiment of the present application is: setting an integrated data acquisition network comprising a collection unit group and an intelligent node, performing athlete identity matching and binding on the collection unit group, collecting image data and athlete sensing basic data group in real time within a motion capture period, evaluating the confidence state of the collection unit in the collection unit group, if the confidence state is credible, calculating high-precision three-dimensional posture based on the athlete sensing basic data group, otherwise, using inverse kinematics algorithm for posture completion, integrating to generate IMU calculation information, obtaining image data to judge the occlusion state of the athlete in real time, marking the occlusion period, dynamically selecting and executing the corresponding data processing strategy according to whether the athlete is in the occlusion period, combining the image data and the IMU calculation information to generate the optimal posture estimation of the athlete, and arranging to obtain a complete posture action sequence.

[0078] The basic principles, main features and advantages of the present application are shown and described above. Those skilled in the art should understand that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A motion capture method based on sports training, characterized in that: include: An integrated data acquisition network consisting of acquisition unit groups and intelligent nodes is set up. The acquisition unit groups are matched and bound to the athletes' identities. During the motion capture period, image data and the athletes' basic sensor data groups are collected in real time. The confidence status of the acquisition units within the acquisition unit group is evaluated. If the confidence status is reliable, the high-precision three-dimensional posture is calculated based on the athlete's sensor data set. Otherwise, the inverse kinematics algorithm is used to complete the posture and integrate the data to generate IMU solution information. Acquire image data to determine the athlete's occlusion status in real time and mark the occlusion period; Based on whether it is during the occlusion period, the corresponding data processing strategy is dynamically selected and executed. The optimal posture estimate of the athlete is generated by combining image data and IMU solution information, and the complete posture and movement sequence is obtained. The specific method for matching and binding athlete identities is as follows: For each athlete, construct an identity feature database containing appearance feature vectors; The integrated data acquisition network collects inertial data, coordinate data, and image data. The coordinate data includes the three-dimensional position coordinates of the acquisition units in the acquisition unit group in the global coordinate system. The image data includes the node field-of-view images captured by the high-definition camera of each intelligent node. Based on the coordinate data of the acquisition unit group, candidate search regions are generated on the corresponding node field-of-view images. Within the candidate search regions, the target detection algorithm is run to locate the pixel-level bounding box of the athlete. A feature extraction model is constructed and used to extract the appearance feature vector and match it with the identity feature library. The cosine similarity is calculated as the matching metric. The optimal allocation algorithm is applied to achieve athlete identity matching and binding based on the principle of the highest matching metric. Data processing strategies include tightly coupled fusion strategies, specifically: Based on the nodal vision images of the athlete in the image data, the three-dimensional coordinates of the athlete's key joints are calculated by binocular visual triangulation as visual solution information. The Kalman filter algorithm is used to accurately calibrate the long-term error caused by sensor drift accumulation in the IMU solution information with visual solution information, and output the optimal posture estimate of the athlete at the current moment. The data processing strategy also includes an estimation backtracking correction strategy, specifically as follows: During the occlusion period, the athlete's initial posture is estimated. At the end of the occlusion period, the athlete's initial posture estimates during the occlusion period are organized into an occlusion posture estimation sequence according to the time sequence. An optimization problem is constructed by combining the occlusion posture estimation sequence with the visual solution information at the end of the occlusion period. The optimization problem takes minimizing the visual reprojection error and posture smoothness as the joint objective function. Batch optimization and adjustment are performed by nonlinear least squares optimization algorithm to obtain the athlete's optimal posture estimate during the occlusion period.

2. The motion capture method based on sports training according to claim 1, characterized in that: The candidate search region is generated as follows: The coordinate data of the acquisition unit group is acquired. The coordinate data is projected onto the two-dimensional image plane of the field of view image of the corresponding high-definition camera acquisition node through the pre-calibrated camera extrinsic parameter matrix of the high-definition camera used, so as to obtain the corresponding pixel coordinate points. Combined with the preset athlete body proportion parameters, a candidate search area of ​​fixed size is generated in the node field of view image.

3. The motion capture method based on sports training according to claim 1, characterized in that: The confidence level is assessed as follows: Based on a predefined human skeletal model and biomechanical constraints, reasonable physiological ranges are set for the relative distances and angles between acquisition units within each limb segment. The confidence state of the acquisition unit is initialized and marked as trustworthy. The relative positional relationship between each acquisition unit is continuously monitored. When the relative positional relationship continuously exceeds the reasonable physiological range within the preset upper limit of deviation time, the confidence state of the corresponding acquisition unit is marked as untrustworthy.

4. The motion capture method based on sports training according to claim 1, characterized in that: The method for calculating high-precision 3D attitude is as follows: The athlete's basic sensor data set includes inertial data and coordinate data collected by the corresponding acquisition unit group. The inertial data includes angular velocity, acceleration and magnetic field vector. A sliding time window is set, and the inertial data within the window is organized into an inertial time sequence and preprocessed. A sensor fusion algorithm based on complementary filtering is used to fuse and solve the inertial data. The high-precision three-dimensional attitude of the acquisition unit in the global coordinate system is output through the quaternion update equation.

5. The motion capture method based on sports training according to claim 4, characterized in that: The posture completion method is as follows: Using a predefined human skeleton model and the high-precision 3D poses of all acquisition units with a trustworthy state within the acquisition unit group, an inverse kinematics solver based on an optimization algorithm is employed. The objective function is to minimize the error between the actual high-precision 3D pose of the acquisition unit and the calculated pose of the corresponding joints in the human skeleton model. The optimal pose of all relevant nodes in the human skeleton model is calculated in real time through numerical optimization methods, and the high-precision 3D pose of acquisition units with an untrustworthy state is deduced and predicted.

6. The motion capture method based on sports training according to claim 1, characterized in that: The method for marking the period of occlusion is as follows: The image data includes the node's field-of-view images captured by the high-definition camera of each smart node; The athlete's coordinate data is projected onto the node's field of view image to generate candidate search regions. A deep learning-based occlusion evaluation model is used to identify the candidate search regions and output the occlusion type, which includes incomplete athlete outline, overlapping occluders, or no occlusion. If all candidate search regions generated for the athlete are identified as no occlusion, the athlete's occlusion status is determined to be no occlusion; otherwise, it is determined to be occluded. The time period when the athlete's occlusion status is occluded is marked as the occlusion time period.

7. The motion capture method based on sports training according to claim 1, characterized in that: The preliminary attitude estimation method is as follows: Using the coordinate data in the athlete's sensor baseline data set as spatial constraints, the athlete's IMU solution information is optimized and solved under a unified inverse kinematics framework to estimate the athlete's preliminary attitude during the occlusion period.

Images