Vision-based Overhead Front Point Recognition of Vehicles for Traffic Safety Analysis
Byeongjoon Noh, Wonjun No, David Lee · 2018
Pedestrian-vehicle accidents are the cause of many human injuries and deaths. To address this challenge, vision-based traffic systems have focused on detecting traffic-related objects' behaviors, such as vehicle position and velocity relative to pedestrians. In this paper, we propose a new and simple model for effectively recognizing overhead front point of vehicles, while only using a single stationary camera capturing from an oblique angle. The proposed system uses faster R-CNN model for detecting object bounding box and mask, projects the mask's extreme points down to find the car's ground front point, and transforms these coordinates from oblique to overhead frame of reference. Our experimental result shows that this method is effective for recognizing overhead front point of car (accuracy: 92.4%) within a certain tolerance.