Research And Application Of Human Motion Behavior Recognition Methods Based On AlphaPose And LSTM

In recent years,people’s enthusiasm for human action recognition research has been rising,and relevant technical achievements have been applied to various fields,such as video surveillance,human-computer interaction,VR entertainment,transportation,etc.In early August 2021,the State Council released the National Fitness Plan(2021-2025),which mentions the use of technology to achieve intelligent sports.In terms of behavior recognition,methods based on RGB video or image information have low detection accuracy and generalization ability;the use of Kinect camera to detect human skeleton information for characterizing movement is more advantageous,but more dependent on external device conditions;currently most means of action recognition are based on the global features of a single picture,but the action is temporal in nature and there is similarity between different movements,so only recognition of current human action from a single fixed point is not comprehensive.To sum up,combining human behavior recognition technology to achieve intelligent fitness goals,as well as to effectively circumvent the above problems,this thesis proposes the use of image information to obtain human posture data and combine it with an action classification network designed based on LSTM to build a motion recognition model and achieve intelligent monitoring of sports and fitness behaviors.AlphaPose algorithm is a popular and more accurate algorithm for multi-person pose detection.First,this thesis investigates the main ways of AlphaPose algorithm to solve the redundant pose problem and improve the accuracy;then,three YOLO family target detection algorithms,YOLOv3-SPP,YOLOv3-tiny and YOLOv5 n,are used to implement the target detection task and are compared in terms of speed and accuracy.The target detection algorithm with the best results among the three is selected to achieve human joint point detection with the AlphaPose backbone structure;after that,the form of the feature data of the action is designed for the 26 skeletal point coordinates of the human body obtained by the AlphaPose algorithm,and the feature terms are 15 vector coordinate values of the limbs and torso relative to the neck joint point and 12 joint angle values,and the total angular acceleration value is calculated as the judgment item of whether to choose the current frame or not.Then,based on the form of the motion features and the movement videos collected in advance,the HIIT behavior dataset is constructed for the final design goal of this thesis,which contains 14 types of HIIT actions,considering three influencing factors: indoor and outdoor background,lighting conditions and multiple viewpoints,taking into account the research value and practical value.The motion feature data corresponding to consecutive multi-frame images are immediately input into the classifier for temporal feature extraction and complete classification.The neural network classifier includes: BLSTM network for motion recognition;BLSTM network incorporating the attention mechanism;Double Lstm Mixed Attention neural network designed and implemented to solve the problem of the specific conditions and target design of this topic,incorporating a parallel motion classification model.The HIIT dataset is used for training and comparison experiments of the network models,and the best action recognition model with the best recognition accuracy is selected using the test accuracy as the criterion.Finally,the Py Qt interface development framework is used to design the action recognition test tool.All related algorithms and technologies are unified,the operation and display interface is designed and implemented,and several video recognition tests are performed to prove the feasibility of the overall design of this method and achieve the goal of the project.