Research On Viewpoint Selection Algorithm For Human Action Evaluation

Rehabilitation therapy can effectively solve the sequelae of serious diseases and help patients recover their physiological functions.Rehabilitation training among them is the most common way,with great significance in postoperative recovery of patients.With the popularity of rehabilitation training,the one-on-one rehabilitation training guided by professional therapists in medical institutions can no longer meet the social needs under the current situation.Self-guided home-based rehabilitation training has been recognized by more patients,but the effect of rehabilitation treatment is difficult to be guaranteed due to the lack of supervision and guidance from professional therapists.The above situation makes the demand for intelligent rehabilitation training system in the medical field increasingly prominent.Among the many techniques of intelligent rehabilitation training system,action evaluation plays an important role as a therapist and has the ability to evaluate and improve the quality of human action,which is the main research content of the system.The research tasks of action evaluation are mainly divided into quantification and feedback.The existing researches often focus on quantification of action,ignoring the improvement ability of action feedback on human action quality,thus affecting the effect of rehabilitation training.Visualization of action feedback is the main way to realize human-computer interaction in action evaluation.However,the imaging principle of human visual system often causes occlusion problem in the feedback visual view,which affects the improvement effect of action evaluation on non-standard actions.In order to improve the movement of rehabilitation training,personalized movement feedback is regarded as the research focus of this thesis,and personalized movement evaluation method based on space-time optimization is proposed.Firstly,stacked hourglass network and semi-dynamic hypergraph neural network are used to estimate the three-dimensional human pose and obtain the human skeleton sequence data in the video.Then,the temporal and spatial characteristics of standard and training skeletal sequences are analyzed,and the temporal and spatial alignment of human skeletal sequences is realized by skeletal redirection and improved DTW algorithm,and the joint score of human body is calculated to obtain the quantified result of action evaluation.Finally,the joint score of action quantification is used to calculate the Jaccard similarity of body parts,and the fitting error between two frames is used to construct the objective function.Meanwhile,the action feedback result of action evaluation is obtained by using the skeleton fitting method,and the visual guided action is generated.The experimental results show that the proposed method can make full use of the dynamic temporal and spatial characteristics of human movements in the quantitative task and feedback task of action evaluation to obtain action evaluation more in line with the needs of rehabilitation training,and further help patients improve the quality of action.Aiming at the occlusion problem in of visual view of action feedback,this thesis focuses on the ability to obtain action information of viewpoint,and proposes an optimal viewpoint selection algorithm for personalized action evaluation.Firstly,according to the imaging principle of human vision system,the characteristics of orthographic skeleton views under different viewpoints are analyzed,and two viewpoints selection indexes are proposed to measure the skeleton visibility.Then,according to the depth perception ability of human visual system,the action characteristics of the feedback results of personalized action evaluation are analyzed,and a viewpoint selection index is proposed to measure the significance of action.Finally,the viewpoint selection objective function is constructed,and the improved LM algorithm is used to calculate the optimal viewpoint of the guidance action.The experimental results show that the proposed algorithm can quickly obtain observation points that are consistent with user habits by iteratively optimizing the view "value",and improve the occlusion phenomenon of action feedback visual view.