Research On Human Behavior Recognition Technology Based On Deep Learning

The growth rate of the elderly population in China is accelerating,and because of their weak physique,the elderly often need the care of their families in their daily life,so they need the joint efforts of the whole society.With the development of security and communication technology,video surveillance has become one of the important means to ensure the life of the elderly,and the camera can be used instead of manual viewing.However,traditional surveillance cameras need to invest a lot of manpower to identify small-probability accidents,so this paper attempts to automatically identify the abnormal behavior of the elderly in video surveillance to improve the utilization of surveillance video.to provide more comprehensive protection for the life safety of the elderly.According to the requirements of the identified tasks,the overall scheme of the system is determined.The whole system mainly includes video surveillance preprocessing,human target detection and target tracking and abnormal behavior recognition.Finally,the system interface is designed for human-computer interaction.The video preprocessing method mainly obtains video frames by interval sparse sampling,which can reduce redundancy and improve efficiency;through the use of contrast enhancement and automatic white balance technology,the image quality is improved;and the cache is dynamically updated so that the model can take into account the previous information.In order to reduce the background interference,the target detection method is used.The relevant theories and principles of the classical target detection algorithm are analyzed and selected,and then according to the needs of this topic,the selected YOLOv3 is improved to train a single-class detector with an average accuracy of 59.6%.Considering the frequent occurrence of multi-targets in video surveillance,a target tracking algorithm is adopted to deal with this scene,which can determine the target identity for different human bodies.Based on the analysis of the common abnormal behavior of the elderly,the models of abnormal behavior such as fall,sitting and lying down are defined.Then two schemes are used to study the behavior recognition.The method based on appearance feature classifies the key nodes of human posture model recognition,and selects the relative position features and velocity features of the main key nodes as supplementary conditions for judgment,and achieves a good recognition effect,but the recognition effect is poor for some behavior categories.Therefore,the motion recognition based on spatio-temporal features uses a mixed two-dimensional and three-dimensional convolution model,which achieves high recognition accuracy and real-time performance.The interface development tool Tkinter is used to design the interactive interface,and at the same time,it is designed for the case of multi-channel video stream input.In order to simulate the actual scene,the experimenter used the surveillance camera to shoot in a laboratory area according to the conditions of falling,sitting,lying,walking,standing,standing,sitting and multi-objectives.the system can effectively identify human behavior,and has a good real-time performance.