Research On Abnormal Behavior Recognition Algorithm For The Elderly People In Indoor Surveillance System

As the rate of the aging population rise quickly in the world,the aging problem of China is prominent.Due to the aging of the elderly people physiological functions,there are some physiological diseases.It is very easy to fall down at home alone due to inadvertent walking or sudden illness.Therefore,this thesis studies how to judge whether the elderly have fallen or not under the indoor monitoring system,and then the information can be fed back to the elderly family members or medical workers through the application or the medical platform,so that the elderly can receive timely assistance.The video detection of the behavior of the elderly in the indoor environment mainly uses the camera to capture the image of the elderly,detects and tracks the movement of the elderly,obtains parameters of the behavior information of the elderly and corresponding state of motion,and finally judge that the status of the elderly in the indoor activity is normal activities or not.The study is mainly divided into three steps,the detection of moving human body,the tracking of moving human body and classification of moving human body behavior.In this thesis,normal walking,sitting and so on are defined as the normal behavior of the elderly,and the behaviors such as tilting,falling,and squatting are defined as abnormal behaviors of the elderly.The research content of this paper are as follows.1.Moving human body detection,this thesis improves the moving human body detection algorithm by establishing a double background model and integrating visual background extractor algorithm with the YUV color space.The moving human body detection method improved first optimizes the sample sampling in the process of sample sampling in this thesis,reduces the resampling rate,and then removes the shadow of traditional Vi Be algorithm in the foreground by utilizing the characteristics of luminance and chrominance separation in the YUV color space model.After that improve the accuracy of the classification of human behavior,and finally obtain the target of the moving human body through segmentation.2.Moving human body tracking,In order to achieve the real-time tracking of indoor human activities,this thesis analyzes the original Kalman filtering algorithm and the Mean Shift algorithm the problems of tracking the moving human body,and combines the advantages and disadvantages of the two algorithms to design a kind of fusion algorithm that combines Kalman filtering and Mean Shift.Through a large number of experiments on single and multiple targets,we can conclude that combining the two algorithms can better track the situation of the elderly in the real-time activities.3.Classification of moving human body,this paper propose a six-layer convolutional neural network model to improve the behavior of the elderly by falling over the classic Le Net model.In order to more clearly show the convolutional neural networks how to extract human features,this thesis introduces a method of Image feature visualization based on deconvolution.In the feature visualization experiment,the color image of the original human body and the binarized human body image after the shadow removal are respectively input,and the human body feature images are extracted and displayed by each layer.Experiments show that the improved convolutional neural network model can effectively distinguish the normal behavior and abnormal behavior of the elderly in indoor environment,and its recognition accuracy reaches 96.15%,which satisfies requirements for accurate recognition of moving human body.