Detecting Avoidance Behaviors of Crowd by Witnessing of Incidents

Abstract

The Tokyo Olympic was held in 2021. Therefore, in Japan, there was an increasing demand for anti-terrorism measures and safety enhancement using ICT. The police have installed surveillance cameras on the streets for 24 hours in order to prevent crimes in the downtown area and to prevent damage from occurring. To arrest the criminals, they sometimes ask the police to use the images from the security cameras installed in shops or individuals to help them in their investigations. However, there are still some problems in the human monitoring system of missing and responsiveness. The purpose of this study is to detect the occurrence of an accident by measuring the behavior of the crowd captured by a security camera, not the suspicious person. Recently, the detection of criminal behavior by learning human behavior using machine learning has attracting attention, but it considered difficult to detect actions due to human overlap. Therefore, we set the assumption that the crowd who observed the incident would take evasive action. The behaviors of the crowd is estimated from the camera and the occurrence of the incident is detected. In the experiment, the behavior of crowd was reproduced using a simulation, and evasive action was detected from the trajectory of the behavior. Evasive action could be detected with 68% accuracy in the most crowded situations.