An Absolute Distance Estimation Method for Obstacle Detection Based on YOLOv5s
Wenyan Ci, Lu Tian, Yangxun Ge, Hongyi Hou, Jihua Ma · 2024
Obstacle detection and distance estimation are key tasks in the implementation of Advanced Driving Assistance Systems (ADAS). This paper proposes a method for estimating the absolute distance between a monocular camera and an obstacle based on YOLOv5 (You Only Look Once) object detection. Firstly, the accuracy of monocular distance estimation relies on the precision of the bounding boxes obtained after object detection. A bounding box post-processing mechanism is proposed to improve the precision of the bounding boxes used in distance estimation. Then, the accuracy of relative distance estimation depends on the selection of key points on the bounding box and the prior knowledge of obstacle height. A method for converting relative distance to absolute distance is introduced to reduce the dependency on these factors. The proposed method for obstacle detection and distance estimation is evaluated using the KITTI dataset. The results demonstrate that within a range of 40 meters, the obstacle detection rate reaches 90%, and the distance estimation error is kept within 3%.