Anticipation of Traffic Accident by Artificial Neural Network Model

Abstract

The most important factor of any vehicle is to drive safely. This is the same with autonomous driving technology. For safe driving, it should anticipate and prevent traffic accident that may happen suddenly in complex traffic environments. In order to avoid accidents, one way is to anticipate which objects will cause accidents and then avoid properly them. Recently, deep neural network technology has been applied to studies to anticipate traffic accidents. These accident anticipation methods showed high performance in accident anticipation task. However, the accident object anticipation task that can say which object is going to collide showed relatively low performance. This thesis proposes a new deep learning based method to improve accident object anticipation performance using only the 1st person camera in a vehicle. Among the various types of traffic accidents, this study focuses on the ego-involved accident because drivers must avoid his/her accidents. The main idea is based on the fact that “human drivers predict future locations and movement of other objects and anticipate whether some objects are a threat to his/her ego-vehicle”. The proposed deep neural network model can anticipate the ego-involved accidents only with the predicted future location and motion of other objects by the probability of accidents for each object in the frame. Effectiveness of the proposed method was validated by testing the dataset that includes multiple accident videos and normal driving videos. The performance was compared with the previous accident anticipation methods, showing better performance in accident object anticipation than the previous accident anticipation methods.