Dept. of Robotics Science and Technology,
Chubu University

Deep Learning Conference

Potential Risk Estimation with Single Monocular Camera

Author
Kota Shimomura, Hiroki Adachi, Tsubasa Hirakawa, Takayoshi Yamashita, Hironobu Fujiyoshi, Masamitsu Tsuchiya, Yuji Yasui
Publication
Secure and Safe Autonomous Driving Workshop and Challenge on CVPR 2023

Download: PDF (English)

Object detection, segmentation, position estimation, and identification of white lines on roads are essential components of computer vision for recognizing surrounding vehicles and pedestrians.
These tasks are focused on differentiating objects and driving scenes with the help of cameras and LiDAR sensors.
However, to enhance the capability of autonomous driving, it is essential to address the possibility of future risks and object variations, which have not been adequately explored.
Specifically, identifying the zones where pedestrians and vehicles may suddenly appear is of paramount importance for ensuring driving safety and preventing traffic accidents.
In this paper, we propose a novel task that aims to estimate the potential risk regions that can cause traffic accidents.
Our focus is on assessing the risk regions from images taken by an in-vehicle camera installed at the front of the vehicle.
We define a risk region as an area where pedestrians or vehicles may appear, and we annotate the Cityscapes dataset with risk region annotations.
Additionally, we propose an end-to-end network and evaluation metrics for estimating the baseline risk regions.
Our results demonstrate that our approach performs exceptionally well in estimating potential risk regions in various scenarios.
This research is expected to facilitate the establishment of safety tasks in the driving environment and enable autonomous driving systems to identify potential risk regions and drive safely.

Previous Next