Research for Roadside Inspection Support by Recognizing Feature from Video Images of Driving Recorders Using Deep Learning

Abstract

Road managers conduct daily patrol inspections, and since the inspections are mainly conducted visually from inside the vehicle, oversights during the inspections are likely to occur. In order to solve this problem, the authors have been conducting research to improve the efficiency of daily inspection using a drive recorder and automatic road object recognition technology based on deep learning. However, there are two problems: one is that the recognition results of the same road object are output from multiple frames, and the other is that objects with similar shapes are misrecognized. In this study, the authors investigated a method of constructing a recognition model that can accurately recognize similar road objects from video images, and obtain recognition results at the closest approach within a shooting distance where the entire object fits within the angle of view. It was found that the problem could be solved by constructing a recognition model that uses images suitable for daily patrol inspections as training data. In addition, it was found that a policy of grouping road objects with the same tendency is appropriate.