Development of Performance Improvement Method of Deep Learning based Automatic Road Dangerous Object Detection System

Dangerous road objects cause traffic accidents, which makes it imperative to find and remove them quickly. For this purpose, MOLIT(Ministry of Land, Infrastructure, Transportation) is providing a smartphone app-based reporting service for dangerous road objects and is developing an automatic object recognition and classification system. The development project has several phases, and elaborate efforts have been put into every phase to improve its performance. A systematic method to improve the system performance was proposed in this study. The specific process was proposed, and new solutions(neural network, data manipulation) were found through factor analysis. The existing neural network, YOLOv3, was replaced by the better-performing neural network model, YOLOv5. Some reported images were difficult to designate as problems and were excluded, and high-quality images were also added from an open-source image dataset. The solutions were applied to the system, and its performance was validated. Road facilities and road kills were selected. The results showed that the performance improved the detection rates of road facilities and road kills from 71% to 91% and from 67% to 97%, respectively. The proposed method is expected to improve the system performance systematically.