車載カメラ映像に対するヒヤリハット発生対象の特定手法の検討

抄録

In this paper, we tackle to identify near-miss traffic incidents, which is close to traffic accidents, on vehicle camera video recorded by event data recorder. We study three types of identification method: typical video feature descriptor, feature extraction extraction based on DNN composed from CNN and RNN, and combination them. From the results of quantitative evaluation using actual event recorder data, we clarify the effectiveness with each video feature extraction technique.