Pedestrian near-miss analysis on vehicle-mounted driving recorders

Abstract

Recently, a demand for video analysis on vehicle-mounted driving recorders has been increasing in vision-based safety systems, such as for autonomous vehicles. The technology must be positioned one of the most important task, however, the conventional traffic datasets (e.g. KITTI, Caltech Pedestrian) are not included any dangerous scenes (near-miss scenes), even though the objective of a safety system is to avoid danger. In this paper, (i) we create a pedestrian near-miss dataset on vehicle-mounted driving recorders and (ii) propose a method to jointly learns to predict pedestrian detection and its danger level {high, low, no-danger} with convolutional neural networks (CNN) based on the ResNets. According to the result, we demonstrate the effectiveness of our approach that achieved 68% accuracy of joint pedestrian detection and danger label prediction, and 58.6fps processing time on the self-collected pedestrian near-miss dataset.