Abstract Summary
There were many cases of people falling and being seriously injured in the past earthquakes. It is feared that many human casualties will occur in the earthquake that is expected to occur in the near future. In order to reduce human damage, it is important to quickly ascertain whether or not there are injured people. In this study, we propose a method to automatically detect human falls during an earthquake. Furthermore, in order to verify the validity of the proposed method, we performed an indoor earthquake simulation using a physics engine. By applying the proposed method to videos generated by simulation, the possibility of detecting falls of the human body during earthquake was investigated. In order to detect the fall of human, it is assumed that the situation in the room is being filmed by a video camera. The behavior of the human body in the video was detected by a neural network for human skeleton detection using the deep learning framework. The camera is considered as a point light source, and human shadow is projected onto the floor. The shadow length of the human body, which is the distance between the point where the straight line connecting the camera and the human head intersects with the floor and the human foot, is calculated based on the coordinates of human skeleton evaluated by the neural network. If the camera is far enough away from the human body, the shadow length of the human body will be longer than the tall of the human body. When the human falls, the shadow length of the human body matches the height of the human. From these facts, we consider the camera as a point light source, compare the shadow length and the tall of the human body, and if the two match, we can determine that the human has fallen. The indoor model was constructed using the physics engine. Furniture models such as bookshelves and cupboards were placed along the walls of the room, and a human body model was placed in the center of the room. Then the strong motion record observed during the 1995 Hyogo-ken Nanbu earthquake occurred in Japan was input to the floor. When the human body shook with the shaking of the earthquake, the shadow length of the human also changed. When the human fell, the length of the shadow suddenly shortened, and the shadow length became almost equal to the tall of the human body. This means that the human had fallen. It was shown that the method proposed in this study can automatically determine human falls and injuries during an earthquake.
