A Novel Pedestrian Positioning System Using Monocular in-Vehicle Cameras
Huaqing Shao, Jiuchao Qian, Xiaoguang Zhu, Peilin Liu · 2019
Monocular depth estimation and pedestrian positioning have become promising in recent years. However, there are few systems specifically designed for pedestrian positioning and most of the current depth estimators deal with the whole images which is redundant. We propose a pedestrian positioning system implemented on monocular in-vehicle cameras. This is an end-to-end system integrating pedestrian detection algorithm, distance estimation algorithm and mapping algorithm. The input of the system is monocular images, and the output which can be visualized is the position information of pedestrians. In order to improve efficiency of the system, we propose a depth estimation algorithm using bounding boxes to obtain distances of pedestrians. Besides, a new pedestrian occlusion (NPO) dataset is created and an anti-occlusion detection algorithm is proposed based on the NPO dataset to eliminate the effects of pedestrian occlusion. In contrast to the current depth estimators, our new system can obtain the whole GNSS information of pedestrians rather than the depth information only. Our method produces competitive results for monocular depth estimation on the KITTI driving dataset. Furthermore, on the dataset collected by ourselves, the system outperforms the current depth estimators especially for estimating long distances.