Research On Visual Action Recognition And Application Of Sedentary Intervention For The Elderly

The aging phenomenon in China is becoming more and more serious,it is very important to improve the healthy pension industry for the development of society.At present,sedentary behavior is a great threat to the health of the elderly,which is the cause of many chronic diseases.Many studies show that periodic exercise is the best intervention for sedentary behavior.However,many elderly people can’t get the supervision and guidance of professional medical staff,resulting in the increase of the mortality rate of many chronic diseases.In view of the above needs,this paper focuses on the recognition of sedentary intervention action based on monocular vision.Through the mobile robot loaded with monocular camera,it can assist the elderly in sedentary intervention without the on-site supervision and guidance of medical staff.Periodic exercise is the best intervention for sedentary behavior.In this paper,an action recognition algorithm based on posture sequence is designed to recognize five kinds of sedentary intervention actions formulated by medical staff through the information transmitted by monocular camera.Firstly,openpose is used to detect 182 D key points representing human body in the video,14 of which are selected,and the pose sequence composed of key points is preprocessed by normalization processing,median filling missing key points and other methods;then,based on the structure of dual stream network,the vectors between the key points in the same frame are numbered by the vector number The angle feature defined by the cosine value of the angle between two frames is input into the space flow of the dual flow classification network,and the feature is learned by the three-layer independent cyclic network in the space flow.The vector difference of the coordinates of the same bone center point in the adjacent frame image is input into the motion flow of the dual flow network as the feature vector,and the feature is learned by the four-layer 1D convolutional neural network in the motion flow the prediction probability of each action output by the network is fused according to a certain weight,and the action type is identified in the final decision layer.In order to make the mobile robot follow the elderly who need assistance better,this paper proposes a target tracking algorithm based on multi feature fusion of target detection.Firstly,the method based on object detection is used to detect the elderly from RGB video;secondly,the motion features of the detection results are extracted by Kalman filter to solve the problem of target occlusion,and the appearance features of the detection results are extracted by deep learning to solve the problem of tracking disorder,and the similarity of the tracking sequence is calculated by weighted fusion of the two features;finally,the tracking sequence with similar similarity is obtained The Hungarian algorithm is used for association matching operation to achieve stable visual tracking effect.In the elderly sedentary intervention assistant system,this paper controls the robot’s forward and backward through the change of the distance between the elderly and the monocular camera.Using the information of the tracking frame and the target frame,a method based on the intersection ratio of two frames to accelerate and decelerate is proposed to improve the stability of the robot following the target.The action recognition module of the system selects the optimal sliding window through the soft NMS algorithm to improve the stability of online recognition of sedentary intervention.