Traffic Risk Estimation from On-vehicle Video by Region-based Spatio-temporal DNN trained using Comparative Loss

抄録

We propose a method to estimate the traffic risk during road navigation based on the region-based spatio-temporal deep neural network (DNN) trained by the comparative loss function. In this method, moving object regions are extracted using the object detector YOLO and their features are clipped out from the middle layer of the detector. Then, these feature sequence is used to estimate the traffic risk by the spatio-temporal DNN followed by the risk estimation network. Experiments were conducted using the KITTI and Dashcam Accident dataset images and the proposed method achieved a positive result using Gaussian noise for robust training. From the results, we have shown that it is possible to estimate a dangerous traffic situation not only caused by accidents but also triggered by congestion using the proposed risk estimation network.