Research On Human Fall Detection Algorithm In Indoor Environment

With the development of deep learning,people try to use the technology of deep learning to detect human fall behavior,so as to find out in time when the fall behavior occurs and send it to the hospital for medical treatment,so as to reduce the harm caused by the fall behavior to the human body.At present,the detection of falls in indoor environment is mostly to detect the target itself,and the accuracy is low and the speed is slow.Therefore,this paper focuses on the fall detection task in the indoor environment,researches the fall detection method based on computer vision,and improves the object detection,feature extraction,and the final behavior classification,so that the fall detection algorithm in the indoor environment can improve the detection accuracy while detecting in real time.The achievements of this paper include the following aspects:First,in the aspect of target detection,the Faster RCNN target detection algorithm has the problems of low network operation speed and low accuracy,so its algorithm framework and algorithm flow are improved.In this algorithm,Res Net is used as the backbone network,and the moving target is predetermined by the inter frame difference method,and then the images in the moving target detection frame and its surrounding extension frame are used as candidate boxes and then input to Faster RCNN,which reduces the number of RPN pre selection boxes generated;At the same time,the feature is extracted layer by layer by referring to the feature pyramid structure,and the deep and shallow features are weighted and fused to improve the detection accuracy,ensuring that the human and non-human moving objects can be accurately distinguished in real time when moving objects appear.On this basis,in the aspect of target tracking,Deepsort algorithm is selected as the target tracking algorithm,and the improved Faster RCNN target detection algorithm is combined with Deepsort algorithm.By associating and matching the Faster RCNN target detection algorithm detection box with the track generated by Deepsort algorithm as the input,the tracking box of Deepsort algorithm is calibrated,thus improving the tracking accuracy of the target tracking algorithm,It ensures that there will be no missing detection and ID replacement when the target is blocked.Secondly,in the aspect of behavior classification,a fall detection model based on support vector machine is proposed from the perspective of video itself.This algorithm uses the above target tracking algorithm to extract the human body’s motion trajectory,and C3 D network instead of C2 D network to extract the characteristics of the motion trajectory.Then,the SVM classifier is improved by combining genetic algorithm to detect falls.This model realizes the image feature fusion at equal intervals.In addition,from the perspective that human beings have two visual information processing channels,a fall detection algorithm based on multi feature fusion is proposed.This algorithm also uses the above target tracking algorithm to extract the human motion trajectory,and then uses Inception V2 network to extract features of different dimensions in the human motion process.Features of different latitudes are weighted and fused at the decision-making level.Finally,Soft max algorithm is used for feature classification,Compared with the fall detection model of support vector machine,the multi feature fusion mechanism based on decision level is implemented,which further improves the accuracy of the algorithm,but the speed of the algorithm is reduced,so the algorithm is sui Table for relatively complex indoor environments.