Research On Fall Detection Method In Bus Compartment Scene

In recent years,public transport dangerous incidents have occurred frequently,and the public transport safety problems reflected in them have attracted more and more attention of the whole society.When the bus suddenly loses control for some reason or other emergencies occur in the bus compartment,it is often accompanied by the fall of passengers.Therefore,using the video surveillance system to alarm the fall behavior of people in the bus compartment can effectively ensure the safety of vehicles and people,and can also prevent more serious consequences.Based on this,combined with technologies such as image processing,deep learning and video analysis,this thesis focuses on the fall detection method in the bus compartment scene with the support of the self-developed intelligent video surveillance system.The main work contents are as follows:(1)According to the scene characteristics of the bus compartment,the bus intelligent video surveillance system is constructed.Taking the bus as the platform,the core computing unit for video processing is constructed based on NVIDIA Jetson Xavier NX module.(2)Based on the methods of deep learning,a human detection algorithm combining object detection and pose estimation is proposed.It is applied to the scene of bus compartment,which can solve the problems of serious occlusion and difficulties in obtaining the features of human.In this phase,the object detection algorithm is used to obtain the position of human's bounding box,and the pose estimation algorithm is used to obtain the position of key points.Then,the fusion strategy of the two is designed based on the spatial coincidence degree.During the experiment,this thesis collects and makes its own data sets,including human object data set and human key point data set,which are used for the training and verification of the algorithm.The experimental results of human detection show that the fused human detection algorithm can obtain comprehensive global and local features compared with a single algorithm.(3)Based on the information in the image,a static detection algorithm of fall behavior in the bus compartment scene is proposed,including the setting of fall discrimination conditions and the design of fall discrimination network.The discrimination conditions are based on the aspect ratio and spatial position of the bounding boxes.The discrimination network takes the key points of human as the features,takes the skeleton as the input,and adopts the supervised learning method for training.Combining the discrimination network and discrimination conditions,the result selection strategy is designed to realize the image-based fall detection.The experimental results show that the proposed method improves the accuracy compared with the single discrimination method.(4)Using the information in video,based on multi-object tracking algorithm and combined with appearance features of human,a two-stage object matching algorithm is designed to realize the dynamic detection of fall behavior based on video.Through the two-stage object matching algorithm,the behavior states of human in consecutive multi-frame images are associated,the number of abnormal frames is accumulated,and a fall alarm is sent.During the experiment,the fall video set of people in the bus is photographed and constructed,and the video-based fall detection method is verified.The results show that the fall detection algorithm combined with two-stage object matching can reduce the false alarm rate and the occurrence of system false alarm.