CN-121983238-A - Body-building action recognition and assessment method based on sparse inertial sensor

CN121983238ACN 121983238 ACN121983238 ACN 121983238ACN-121983238-A

Abstract

The invention discloses a body-building action recognition and evaluation method based on a sparse inertial sensor, which belongs to the technical field of artificial intelligence and motion health, wherein motion information of a human body 6 node is obtained through the inertial sensor, motion information of a human body 24 node is obtained through a human body motion reconstruction network, the motion information is input into a trained body-building action classification model, and finally the body-building action is quantitatively evaluated by utilizing Dynamic Time Warping (DTW) according to a body-building action recognition result, so that the real-time body-building action is rapidly compared with a standard action. By adopting the body-building action recognition and evaluation method based on the sparse inertial sensor, the accurate recognition and quantitative evaluation of body-building actions can be effectively realized, and meanwhile, the method is low in cost and higher in practical applicability.

Inventors

LI QIAN
TIAN JIPENG
SUN XUDONG
CHEN YUCHEN
BEN YUEYANG
WU LEI

Assignees

哈尔滨工程大学

Dates

Publication Date: 20260505
Application Date: 20260121

Claims (9)

1. A body-building action recognition and evaluation method based on a sparse inertial sensor is characterized by comprising the following steps: s1, acquiring original motion measurement data of a specific human body node based on a plurality of inertial sensors sparsely arranged on the node; S2, inputting the original motion measurement data into a pre-trained human body motion reconstruction network, and reconstructing to obtain motion information of a plurality of joint nodes of the whole body of the human body; S3, inputting the motion information of the reconstructed joint nodes of the whole body of the human body into a pre-trained body-building action classification model, and identifying to obtain a classification result of the current body-building action; And S4, matching a corresponding standard template from a pre-stored standard action template library according to the classification result, comparing the motion information of the current action with the standard template by utilizing a dynamic time warping algorithm, and outputting a quantitative evaluation result.
2. The method for recognizing and evaluating exercise actions based on sparse inertial sensors according to claim 1, wherein in step S1, 6 inertial sensors are disposed on the head, abdomen, left wrist, right wrist, left calf and right calf nodes of the human body, respectively, and the raw motion measurement data includes at least angular velocity information and acceleration information measured by gyroscopes and accelerometers in the inertial sensors.
3. The sparse inertial sensor-based exercise motion recognition and assessment method according to claim 2, wherein step S1 further comprises filtering the angular velocity information, calculating attitude angle information at each sensor node through a quaternion differential equation, and forming motion information of each node by the acceleration information and the attitude angle information together.
4. The sparse inertial sensor-based fitness action recognition and assessment method according to claim 1, wherein the human motion reconstruction network in step S2 is a neural network based on a bidirectional gating cycle unit GRU; the motion reconstruction step specifically comprises the following steps: Performing calibration and normalization preprocessing on the original motion measurement data, and converting the original motion measurement data into a unified SMPL human body parameter model coordinate system; Inputting the preprocessed data into the human body motion reconstruction network, and predicting to obtain a gesture rotation matrix and motion acceleration of a plurality of joint nodes of the whole human body under the human body parameter model coordinate system.
5. The sparse inertial sensor-based exercise motion recognition and assessment method of claim 4, wherein the calibration process is implemented based on the following mathematical relationship: ; Wherein, the Representing a posture rotation matrix between each inertial sensor coordinate system and a bone coordinate system at a corresponding joint, Represent the first A gesture rotation matrix of rotation relationship between the skeletal coordinate system of the individual joints and the SMPL manikin coordinate system, Represent the first The attitude of the individual inertial sensors themselves rotate the matrix, Representing a transition matrix from the global inertial coordinate system to the SMPL coordinate system.
6. The sparse inertial sensor-based exercise motion recognition and assessment method of claim 5, wherein the normalization preprocessing comprises normalization operations performed on master and slave blade joint measurements, respectively: normalized calculations from leaf joint measurements are: ; Wherein, the 、 Representing normalized slave blade joint measurements, 、 A pose rotation matrix between the joint skeleton coordinate system and the SMPL coordinate system of the master blade and the slave blade respectively; 、 bone motion acceleration of the master and slave leaves in the SMPL coordinate system are represented respectively; the normalization calculation of the primary blade joint measurements is: ; Wherein, the 、 Representing normalized primary blade joint measurements.
7. The method for recognizing and evaluating body-building actions based on sparse inertial sensor according to claim 1, wherein the body-building action classification model in step S3 is an LSTM-RF mixed model composed of a long-short-term memory network LSTM and a random forest RF, and the specific training process comprises: s31, constructing the motion information of the reconstructed joint nodes of the whole body of the human body into an original characteristic data set; s32, inputting the original characteristic data set into the LSTM model, extracting long-term dependency relationship and dynamic change rule in the data, and obtaining output ; S33, outputting the output Fusing the original characteristic data set and constructing a new characteristic data set; s34, inputting the new characteristic data set into the RF model, and finally classifying to obtain a classification result of the body-building action.
8. The method for recognizing and evaluating exercise actions based on sparse inertial sensor according to claim 1, wherein step S4 specifically comprises: S41, selecting a preset key joint according to the classification result; s42, calling a standard action template corresponding to the classification result from a standard template library, wherein the standard action is a standard sequence; S43, extracting time sequence data of the key joint in the motion process in the current action to form a test sequence; S44, calculating a dynamic time warping distance between the test sequence and the standard sequence, and obtaining the quantitative evaluation result based on the distance.
9. The sparse inertial sensor-based exercise motion recognition and assessment method of claim 8, wherein the quantitative assessment result of step S44 is represented by a score, the score being calculated based on the cumulative regular distance obtained by the dynamic time alignment algorithm And is mapped by the following formula: ; Wherein, the The final score is indicated as such, The full-scale of the mark is indicated, The base value is represented by a value of, Representing the impact factor.

Description

Body-building action recognition and assessment method based on sparse inertial sensor Technical Field The invention relates to the technical field of artificial intelligence and sports health, in particular to a body-building action recognition and evaluation method based on a sparse inertial sensor. Background Along with the rapid development of the economy and society and the continuous improvement of the living standard of residents, the health consciousness is increasingly deep, and the attention of the masses to physical exercise and scientific body building is remarkably improved. In this context, how to accurately identify exercise actions, normalize and evaluate effects becomes an important technical requirement in the fields of exercise guidance, exercise rehabilitation and personal health management. Currently, mainstream motion recognition and analysis systems rely on optical sensors to capture and analyze human body gestures through computer vision technology. Although the method can obtain higher recognition precision, the deployment is generally limited by a specific environment, the equipment cost is high, the method is easily interfered by external factors such as light rays and shielding, and the method is difficult to popularize and apply in daily life. To overcome the above limitations, inertial sensor-based motion recognition schemes are becoming increasingly interesting. The inertial sensor has the characteristics of small volume, low power consumption and low cost, and can be conveniently integrated in wearable equipment to realize real-time and continuous acquisition of motion data. However, the existing scheme based on the inertial sensor still has the defects that a part of scheme needs to densely arrange sensors at a plurality of nodes of a human body, so that the wearing is complicated and the user acceptance is low, a part of scheme lacks an efficient motion data complement mechanism, the whole body motion state is difficult to restore through sparse node data, meanwhile, most of evaluation methods can only realize qualitative judgment, cannot accurately and quantitatively evaluate the motion normalization, and cannot meet the scientific body-building requirement of the user. Disclosure of Invention The invention aims to provide a body-building action recognition and assessment method based on a sparse inertial sensor, which keeps the advantages of low cost and strong environmental adaptability of the inertial sensor, solves the problems of complicated wearing, low data complement precision and inaccurate assessment of the existing scheme, and provides technical support for popular scientific body-building. In order to achieve the above purpose, the invention provides a body-building action recognition and evaluation method based on a sparse inertial sensor, which comprises the following steps: s1, acquiring original motion measurement data of a specific human body node based on a plurality of inertial sensors sparsely arranged on the node; S2, inputting the original motion measurement data into a pre-trained human body motion reconstruction network, and reconstructing to obtain motion information of a plurality of joint nodes of the whole body of the human body; S3, inputting the motion information of the reconstructed joint nodes of the whole body of the human body into a pre-trained body-building action classification model, and identifying to obtain a classification result of the current body-building action; And S4, matching a corresponding standard template from a pre-stored standard action template library according to the classification result, comparing the motion information of the current action with the standard template by utilizing a dynamic time warping algorithm, and outputting a quantitative evaluation result. Preferably, in step S1, the inertial sensors are disposed on the head, abdomen, left wrist, right wrist, left calf and right calf nodes of the human body, respectively, and the raw motion measurement data at least includes angular velocity information and acceleration information measured by a gyroscope and an accelerometer in the inertial sensors. Preferably, the step S1 further comprises the steps of filtering the angular velocity information, calculating attitude angle information of each sensor node through a quaternion differential equation, and forming the acceleration information and the attitude angle information into motion information of each node. Preferably, the human motion reconstruction network in step S2 is a neural network based on a bidirectional gating cycle unit GRU; the motion reconstruction step specifically comprises the following steps: Performing calibration and normalization preprocessing on the original motion measurement data, and converting the original motion measurement data into a unified SMPL human body parameter model coordinate system; Deducing the relative positions of the other 5 slave blade joints relative to the maste