Enhanced approach for on-road obstacles detection using level-set-YOLOv3 combination
No Thumbnail Available
Date
2021
Journal Title
Journal ISSN
Volume Title
Publisher
University of Oum El Bouaghi
Abstract
Automatic detection and classification of on-road obstacles is an integral part of Intelligent Transport Systems (ITS). It increases road safety by providing valuable information about the environment where the vehicle navigates using onboard sensors, and can intervene in risky situations. However, it is a difficult task to achieve due to multiple variability conditions presenting in the real urban scenes. These latter can perform partial occlusion, multi-view angles, multi-scale objects, different lighting conditions, etc. In order to solve these problems, we describes in this paper a system for automatic detection and classification of on-road obstacles. The system is achieved by combining the Level-Set segmentation technique and the YOLOv3 algorithm that uses Darknet-53. The YOLOv3 used in our work was pre-trained with COCO dataset, on 80 categories of common objects. We started by fine tuning the model, to detect three types of on-road obstacles: Car, Person, and Bycicle, with KITTI dataset, which contains scenes from public roads. Then we used the segmented RGB images from KITTI, for training and testing the performance of the model. The proposed model with combination of Level-Set and YOLOv3 is compared to a simple YOLOv3 model without any preprocessing. The results
demonstrate the effectiveness of the model with combination in detecting small size, overlapping objects with high precision of
about 87%.
Description
Keywords
ITS; On-road obstacles; Object detection; Object classification; Level-Set; YOLOv3