Traffic Accident Prediction without Object Detection for Single-Vehicle Accidents

抄録

Recently, the research on traffic accident prediction models via deep learning has attracted significant attention. Many recent high-accuracy accident prediction models rely on bounding boxes obtained from object detection, which cannot predict single-vehicle accidents with a high fatality rate because of their structure. This paper proposes a model that predicts single-vehicle accidents by estimating the probability of accident occurrence at the frame level. The proposed model integrates depth and segmentation information along with RGB images and optical flow information to enhance prediction accuracy. To validate the effectiveness of the proposed model in single-vehicle accident scenarios, this study constructed a CARLA Accident Dataset using a driving simulator and a dataset containing only single-vehicle accident scenes selected from the Detection of Traffic Anomaly dataset. The proposed model demonstrated high accuracy in the investigated datasets, indicating its effectiveness in predicting single-vehicle accidents.