Research On Fall Detection Based On Pose Estimation

With the widespread use of video surveillance in daily life,it has become crucial to automatically recognize the movements of the human body in the screen through surveillance.Falling is a factor that threatens people's health.If you can quickly identify and make an early warning,you can protect people's safety in time.As a branch of human behavior understanding,the purpose of fall detection is to identify whether the person in the scene is in a state of fall,therefore fall detection study has important significance.When the human body falls,due to the different reaction mechanisms of each part of the body to danger and the bearing degree of the injury is different,falling actions in different directions will cause different intensities of damage to the human body.Generally,the injury caused by the human body falling forward is smaller than the injury caused by the human body falling backward.Existing fall detection studies usually focus only on whether the conduct occurred to detect a fall,but not concerned about the direction of the problem of falls.According to the above analysis,this paper introduces a fall detection and direction judgment method based on attitude estimation,which solves the problem of human fall detection and fall direction judgment based on attitude estimation for the first time.Firstly,this paper studies the different situations of falls from the perspective of posture analysis,detects the timing of the fall behavior and determines the direction of the fall,so as to make an accurate warning.In order to make accurate fall detection and direction judgment,this paper uses the human body detection method to identify the position of the person in the scene,and then predicts the joint point coordinates of each human body through the posture estimation network.On the basis of obtaining specific feature parameters,the svm classifier is used to classify and train the relevant joint point features to improve the accuracy of fall detection classifier.Then,when the fall detection classifier detects the occurrence of a fall event,it saves the previous and next frame information of the current node and inputs them into the three-dimensional conversion network.The two-dimensional joint point information is converted into three-dimensional joint points through the corresponding algorithm,and all frame sequence parameters are unified.The human body is modeled on the basis of three-dimensional coordinates,and the specific direction of the human body falling is calculated according to the relevant formula.Finally,in order to verify the effectiveness of the fall detection method in this article,we compare experiments with other methods in in Multi Cam,Le2 i and UR datasets.The final experimental results verify that the method proposed in this paper has certain advantages in fall detection tasks.