Research On The Recognition Of Abnormal Human Behaviour Based On Expressions And Body Movements

In recent years,with the increasing awareness of people’s security,most families have installed indoor monitoring equipment,and there are more and more scenarios for intelligent analysis through video monitoring equipment,and identifying abnormal human behavior in indoor security scenarios is a key area.In scenarios such as elderly people living alone,left-behind children and working people living alone,there are abnormal behaviors such as people falling down and having heart attack,which puts a very high test on the real-time,accuracy and robustness of human abnormal behavior recognition algorithm and recognition system.In this paper,we study the problem of human abnormal behavior recognition and real-time monitoring in indoor scenarios such as people living alone.First,this paper introduces the theories and models related to expression and action recognition used in human behavior recognition.The components of common datasets for expression and action recognition and the source,number and class of samples contained in each dataset are introduced.For the problem of few datasets related to specific actions,the fall and heart-covering action datasets are self-made and the quality of the datasets is evaluated.In order that the expression recognition algorithm can perform better feature extraction,an adaptive histogram equalization algorithm is used to adjust the contrast of the expression recognition dataset.Secondly,to ensure the accuracy of face expression recognition,the RepVGG network is used as the basis for face expression recognition algorithm improvement,and the CEN-RepVGG expression recognition algorithm is proposed.The ECANet efficient channel attention module and Center loss loss function are added to the backbone network to improve the model recognition accuracy and adaptability to different recognition scenes.The SGD and Adam optimizer using Nesterov momentum are adopted to jointly optimize the parameters in the model training phase to improve the model performance after the training is completed.The effectiveness of the expression recognition algorithm was validated on the CK+,FER2013 and RAF-DB datasets.Then,the Med-YOLOv5 action recognition algorithm,which combines YOLOv5 and Mediapipe Pose algorithms for the extraction of RBG action information and human skeletal information,respectively,is proposed for the problems of traditional action recognition mostly using skeletal timing features and less,using spatial features and low robustness of recognition accuracy.For the determination of heart position,a heart position detection algorithm based on skeletal point coordinates and a heart-covering action recognition algorithm based on the relationship between skeletal point coordinates and skeletal angle are proposed to judge the heart-covering action by skeletal information.The validation of algorithm effectiveness was carried out on the homemade dataset and compared on different models.Finally,the improved expression recognition and action recognition algorithms are fused,and the human abnormal behavior recognition algorithm based on facial expressions and body movements is proposed,and the expression and action recognition effects are fused at the decision level.The proposed algorithm is validated by building a solitary person abnormal behavior monitoring system.The results show that the proposed algorithm can achieve real-time abnormal behavior recognition and make timely warnings for abnormal behavior.