Image-based fall detection in bus compartment scene

Abstract

Bus, as an important means of public transportation, has always received attention for its safety. When an emergency occurs, fall behaviour is one of representative features to help alarm. However, compared with other open and fixed scenes, the bus compartment has its own characteristics, such as crowded, closed, in moving, and so on. Considering the problems of people blocked as well as light changes, this paper proposes a new image-based fall detection method in bus compartment scene, which can be divided into three parts: (1) Human detection, where object detection and pose estimation algorithms are combined, and then both global and local human features can be obtained; (2) fall discrimination, where both of the fall discrimination conditions and fall discrimination network are designed; (3) alert, where an alarm strategy is designed. Experiments are done in a real bus, and the Nvidia Jetson Xavier NX module is used to analyse the videos and images. Results finally show that the proposed fall detection method in bus compartment scenes can achieve 90% accuracy.