Research On Fall Detection Method Based On Depth Camera And Temporal Network Model Optimizatio

Fall detection is a research hotspot in the field of intelligent video surveillance.Fall will cause serious harm to human body.Nowadays,fall detection method based on vision has become the main research direction.Through the use of image processing,feature model and various network structures,human body tracking,behavior analysis,fall detection and other functions are realized.The RGB image data information is single,the quality of data and the division of behavior become the key factors of model efficiency,and only through the color perception information can not fully represent the characteristics of behavior.In recent years,depth information has become the research focus of fall detection.The depth data in the scene is obtained by machine,and the three-dimensional information of objects in the scene is obtained through calculation.However,due to the large amount of depth information data,higher requirements for accuracy and efficiency are put forward,and the human behavior is more complex,which makes the classification boundary of various behaviors fuzzy,and makes various detection prone to error.In order to solve the above problems,the author has done the following work:(1)This paper proposes a method based on Kinect spatial model.Firstly,Kinect is used to obtain the depth information of each frame and calculate the three-dimensional data of all bone points.In order to locate the human body and design the model,a dynamic coordinate system for real-time tracking the human body is established.When extracting data features,this paper designs and fits three-dimensional formulas and models,and establishes criterion formulas for different key points to calculate their feature mapping.In the detection,this method first designs a specific key point selection algorithm to find the dangerous area of human body,that is,the area with strong characteristics.Methods then use the criterion formula and three-dimensional model to calculate the data estimation value of human body balance,and highly characterize the behavior data,so as to realize stable and rapid detection.The experiment achieves an accuracy of 95.13%,and the detection efficiency is very stable in various scenes.(2)In order to compensate for the dynamic detection performance of the first method,this paper proposes a method based on Kinect speed classification model.Firstly,the continuous depth map is obtained and the speed sequence of human key points is calculated.After analyzing the speed data of various behaviors,the author extracts the data of key points in the key frame as the classification data set,and uses support vector machine one-versus-rest training.The classification model is used as the speed capture channel for detection,which does not need complex data sets and can detect falls quickly and accurately,screen out the cases that do not meet the conditions in speed.When used with the first method,it achieves 96.15% accuracy and solves most error prone situations.(3)This paper presents a fall detection method based on LSTM neural network using information and frame attention module.The importance of various characteristic data of human behavior is different.In order to make full use of effective data characteristics and avoid wasting time on inefficient data,the author designs an information attention module to train the weight of different information,change its utilization rate,and better train the comprehensive characteristics between various data.At the same time,the frame attention module is designed,which uses the temporal network structure to train the frame weight and adaptively train the utilization of different frames.The accuracy rate of this method is94.35%,which fully excavates the value of various data including time,and comprehensively improves the use of various data.