Research On Human Fall Behavior Analysis Technologies Based On Machine Learning

With the rapid development of artificial intelligence and computer vision technology,human behavior analysis has been widely used in many scenarios as an important research direction.Among them,the analysis of human fall behavior can effectively monitor fall events of the elderly or children,and avoid greater damage through timely warning and rescue,which attracts widespread attention in the industry.Based on the home scene,this thesis studies the human fall behavior analysis technologies based on machine learning.The main content of the work is as follows:(1)Relevant algorithms of human fall behavior analysis are researched.Firstly,the foreground object extraction algorithm is introduced.Then,the human fall feature extraction method is studied.Finally,the machine learning classification algorithm is described.(2)An effective extraction algorithm for human posture key points for fall analysis is proposed.Firstly,the Alpha Pose algorithm is used to detect the key points of human skeleton.Then,a hierarchical screening method based on behavior analysis is proposed,which mainly includes a preliminary screening method based on confidence,a middle screening method based on behavior differences,and a final screening method based on behavior consistency.Finally,for some key points that may be missing in some frames,an improved bilinear interpolation(IBLI)theory is proposed to predict and fill the missing key points.(3)A double fall analysis algorithm based on the temporal and spatial features of human posture is proposed.Firstly,a method for extracting temporal and spatial features of human posture for fall analysis is proposed,which mainly includes three parts: a spatial feature design method based on behavior analysis,a timing feature design method based on behavior analysis,and a feature fusion method based on the adaptive frame sliding window(FSW),which is used for extracting the human posture features in consecutive frames within a period of time and fusing them into feature vectors.Then,a dual fall behavior analysis algorithm named FT-SVDD is proposed,that is,the feature threshold is used for fall pre-judgment,and then the SVDD classifier is used for fall re-judgment.Finally,the algorithm proposed in this thesis is tested and verified through experiments,and the results show that the algorithm not only guarantees the speed,but also improves the accuracy of fall detection.