Using deep learning to detect and track all modes in traffic videos

点击翻译按钮获取中文摘要

Transportation network planning, traffic safety and intelligent transportation systems require origin-destination (OD) data to assist in designing and planning for infrastructure such as roadways, intersections and modern roundabouts. OD data is traditionally collected by manually annotating videos of all modes of urban travel (e.g., vehicles, transit, pedestrians and cyclists). Manual annotation is extremely tedious, because one needs to timestamp the entrance and the exit of all modes in a video segment. It requires starting, pausing, rewinding a long video for hours. An intermediate solution has emerged, where vendors send their traffic videos to a software service (such as MioVision) that first processes the videos by a combination of automatic and manual methods, then sends back OD data to the vendors. In this research project, ISL Engineering Services is partnering with Dr. Nilanjan Ray in the Dept. of Computing Science, University of Alberta to utilize deep learning and reinforcement learning to create computer vision algorithms that will completely automate the OD data generation from traffic videos. This research would be a first step toward building a prototype of a completely automated video analysis tool for OD data collection and traffic studies that does not exist in the Canadian market.