Research On Pedestrian Fall Detection Based On Deep Learning

At present,China has entered an aging society.Falling is one of the common factors threatening the physical and mental health of the elderly.With the development of science and technology,researchers try to use more advanced depth convolution technology to detect pedestrian fall behavior,so as to reduce the injury of falls to the elderly as much as possible.Because of the popularization of surveillance equipment such as cameras,the fall detection method based on computer vision has important research significance and value.However,most of the traditional fall detection methods based on computer vision are suitable for the simple scene of individual pedestrians and difficult to be applied to the real scene of multiple pedestrians and complex environment.In order to solve the above problems,we propose a fall detection algorithm based on joint feature and an algorithm based on the combination of pedestrian pose estimating and tracking to capture continuous video frames of each pedestrian target and input them into the YOLOv4-tiny fall detection model.The YOLOv4-tiny fall detection model is deployed in the Jetson Nano hardware terminal to realize fall detection in the real scene.The main work of this thesis is as follows:(1)To solve the problem of slow speed of pedestrian pose estimation due to the large amount of network calculation,this thesis proposes pedestrian pose estimation algorithm based on improved Center Net.The depth separable convolution is used to replace the ordinary convolution of the residual module in Hourglass-104 network to reduce the amount of calculation of the whole network model.Experiments show that the improved Center Net algorithm can improve the speed of pedestrian pose estimation and further reduce the time required for pedestrian fall detection.(2)Aiming at the problem of inaccurate pedestrian tracking caused by pedestrian occlusion in multi pedestrian scene,a pedestrian tracking algorithm based on feature matching is proposed.The pedestrian position is predicted according to the change of pedestrian centroid.By judging whether the pedestrian has overlapping shaded area,it is further determined whether it is necessary to segment the unshaded area as the feature area,extract the pedestrian color features,and select different feature matching methods for tracking according to the different states of the pedestrian.Experiments are carried out in different scenes.The results show that the pedestrian tracking algorithm based on feature matching can better deal with the tracking error caused by pedestrian occlusion and reduce the impact of pedestrian occlusion on the accuracy of fall detection in multi pedestrian scenes.(3)In view of the lack of multi pedestrian fall data set at this stage,this thesis collects multi pedestrian fall image data in many ways,and establishes a multi pedestrian fall data set with 4300 images to train and test the fall model.(4)This thesis proposes a fall detection algorithm based on joint point features and a pedestrian fall detection algorithm based on YOLOv4-tiny.The YOLOv4-tiny model is trained on the self-built multi pedestrian fall data set,and the parameters are adjusted to obtain the fall detection model suitable for the multi pedestrian complex scene and single pedestrian simple scene in this thesis.It is deployed in the Jetson Nano hardware terminal to realize the fall detection in the actual scene.Experiments show that the accuracy of the fall detection algorithm based on YOLOv4-tiny in the public single pedestrian simple scene data set and the self-built multi pedestrian complex scene data set are 95.51% and 91.67% respectively,which is more suitable for the actual scene than the fall detection algorithm based on joint point features,and has certain advantages.