Homography-Based Vehicle Pose Estimation from a Single Image by Using Machine-Learning for Wheel-Region and Tire-Road Contact Point Detection

摘要

Image-based metric measurement and development of traffic surveillance systems have attracted wide interests within academia and industry for the past decade due to recent advancements in computer vision and the processing power required for machine-learning. Utilization of camera vision is gaining attention in this realm, particularly due to its unobtrusiveness. The research objective is to develop an image-based photogrammetry system for measuring vehicle lane pose using a single perspective camera with applications in law enforcement and crash-scene investigation. The proposed algorithm comprises of two steps: (1) Developing a Deep-Learning-based technique for identifying/classifying the wheels on a vehicle, as Regions of Interests (ROI), extracting the tire-road contact point from the image, (2) using a Homography-based approach to extract metric measurements, such as vehicle pose. Our proposed method was tested and evaluated on a large number of images taken at different traffic inspection stations under different lighting conditions and weather differentials to demonstrate its efficiency and robustness. Results are promising. This research can pave the way towards automating the task of flagging truck bypass lanes for law enforcement and also for image-based crash-scene investigation.