Research On Indoor Fall Detection And Behavior Analysis Based On Scene Context

As a frequent and highly injurious abnormal behavior in the elderly,falls not only seriously affect people's physical and mental health,but also bring huge pressure to public health.Manual nursing is effective,but it requires a lot of energy and is very inefficient.With the popularity of surveillance devices,daily activities can be recorded in video.Therefore,video-based behavior analysis has important theoretical research value and social significance.In order to effectively distinguish falls from daily activities,the main work of this paper is as follows.Part one:object detection and feature extraction.Most of the fall detection methods have poor stability of object extraction,which is easily disturbed by motion background and other factors,resulting in poor extraction.In addition,daily activities are complex and varied,and if the extracted features are single or not representative,may can't help get a good result.In this paper,the detection model YOLO v3 was used to detect and locate people at home.After the analysis of daily activities,we extracted the features like shape ratio,movement speed and center height,and then integrated them into a temporal feature sequence for fall detection.Part two:fall identification.Most of the traditional fall detection methods identify falls according to a single image without considering the timing information between frames.This paper proposes a classification method based on convolutional neural network for learning feature sequences and classifying actions.To further improve the ability of the model to obtain global information,statistical features are used to help accurately distinguish the fall behavior from the daily behavior.Part three:context-based fall detection in videos.Since the temporal and spatial information of falls is encoded in a series of images,in order to extract robust features and explore the context information of the scene,a context-based fall detection method is proposed in this paper.By using the three-dimensional convolutional network to model the appearance and movement of the human beings,spatial and temporal features with strong representational ability are extracted to describe the activities.Then,the extracted features are sent to the bidirectional LSTM for high-level feature modeling.By learning the relationship between a series of frames,the accuracy of fall detection can be improved.Experimental results show that the features based on visual cognition can effectively describe the human's activities,and the discriminant method based on convolutional neural network can effectively distinguish the fall from the daily activities on the common data set compared with other fall detection methods.In addition,this paper collected a large number of indoor videos of daily activities for training and testing the context-based fall detection model.The results show that this model can effectively detect the falls,and the end-to-end structure avoids the tedious step design of the traditional method.