Research On Upper Limb Motion Analysis Method Based On Skeleton Point Motion Information

Daily movement analysis of the upper limb has a wide range of applications and important research value in the diagnosis of upper limb neuromuscular system diseases,rehabilitation efficacy assessment,sports science,upper limb exoskeleton design and control,and prosthesis control.Human motion data can be collected by inertial sensors or optical sensors,among which Microsoft’s Kinect can provide 3D spatial coordinates of limb joint points in real time,with the advantages of non-contact,low noise and low cost.Therefore,this paper studies the method of daily movement analysis of upper limbs based on Kinect.The main components include:Seven types of daily movements were selected as the research objects,namely,picking up the cup and placing it at the mouth,putting the cup down from the mouth,moving forward and backward in the horizontal plane,abduction of the upper limb,adduction of the upper limb and folding of the arms.Then,the skeleton data acquisition system of the depth camera was established,and the data acquisition of the upper limb action was completed.The upper limb motion features extracted in this paper include joint velocity,joint acceleration,and the angles,angular velocity,and angular acceleration of shoulder adduction/abduction,shoulder flexion/extension,shoulder internal rotation/external rotation,and elbow flexion/extension.A total of 36 kinematic feature sequences were extracted,and the principal component analysis method was used for feature selection.The dynamic time warping algorithm and bidirectional long short-term memory network were used in the recognition model,and the extracted kinematic features were used as the input of the recognition network.The accuracy rate,precision rate and recall rate were used to verify the recognition results.The accuracy of upper limb action recognition based on dynamic time waring algorithm could reach 78.30%,and the accuracy of upper limb action recognition based on bidirectional long short-term memory network could reach 91.7%,which proved the effectiveness of kinematic features as recognition network features.And bidirectional long short-term memory network is better than the dynamic time warping algorithm.The evaluation model uses the Gaussian mixture algorithm,takes the kinematic features as input,trains the Gaussian mixture model on the normal movement pattern,and quantitatively evaluates the quality of upper limb movement by calculating the likelihood values of the test sequence and the normal model.In this paper,the weighted fusion of the evaluation results on the kinematic feature sequence probability model and the motion index probability model is proposed,and the quantitative score of the upper limb action is obtained as the evaluation result.The evaluation model was validated under different upper limb completion,different body states,and different execution times.After weighted fusion,the Pearson correlation coefficients of the evaluation models were 0.89,0.83 and 0.91,respectively.The evaluation model based on weighted fusion more accurately reflected the individual’s upper limb movement,which proved that the upper limb movement evaluation model proposed in this paper could effectively evaluate the upper limb movement.