A Fall Event Detection Based On Attention Guided Bi-Directional LSTM Network

In recent years,with the rising trend of China's population aging,the issue of aging population has become more and more frequent in people's view,mainly because the demographic structure of our country has changed greatly,with the proportion of elderly population increasing year by year,while the low willingness of young people to have children has led to a decline in the proportion of newborn population year by year.And in the foreseeable future,the rise of the aging population will become a trend,so the solution of this problem becomes especially urgent,and fall events,considered as one of the biggest risks threatening the health of the elderly,are the focus of this hot issue,and because of this,the topic of fall detection frequently appears among maj or research institutions.More and more experts and scholars are involved in the task of fall detection in elderly people living alone scenarios.The traditional fall event detection research is commonly conducted in ideal experimental environment,without considering the challenge of complex background in real situation.Therefore,thi paper aims to detect fall event detection in complex background based on visual data.Different from most conventional background subtraction methods which depend on background modeling,Mask R-CNN method is first used to accurately extract the moving objects in the noise background.Then,an attention guided Bi-directional LSTM model is proposed for the final fall event detection.Based on this idea,our paper mainly including the following contents:(1)Establishing a database of human behavior in complex background.In a laboratory environment,a normal RGB camera was used to record human behavior video clips,including several normal behaviors and falling behaviors in indoor environments,such as:walking,jumping,squatting,and lying down.A total of 60 videos contain 60 falling behavior clips.Finally,we added Gaussian noise to this dataset to approximate the noise generated in the real environment.(2)Background substraction.The self-collected video dataset is first preprocess,and then Mask R-CNN is used to extract ideal human contour from video frames with complex background.In the experimental part,we compared several traditional background subtraction methods to present the effect of Mask R-CNN used in this article.(3)Fall event detection.After processing by Mask R-CNN,VGG-16 is used to extract the features of the binary image,and the output of the last layer is feed to the Attention Guided Bi-directional LSTM network designed in this paper for the final fall behavior detection.In the experimental part,we compared the traditional methods and deep learning methods proposed in recent studies on the self-built dataset to prove that the proposed algorithm has higher accuracy and robustness.The system designed in our paper includes video data collection,background substraction,feature extraction and fall classification modules,which can ensure the robustness and accuracy of system in complex background environment.In order to verify the effectiveness of our method,we compare the proposed method with other methods in the public dataset and self-built dataset.In the experiment,various indicators show that the method we designed has better accuracy and robustness.This means that our method is very suitable for the fall events detection in complex scenes.