Collision Avoidance Approach for Autonomous Driving Using Instance Segmentation

While autonomous driving is a growing trend in research, it is difficult to commercialize widely with a system that requires expensive sensors. In this paper, we only use a low-cost camera sensor to propose an autonomous driving algorithm including collision avoidance. The proposed algorithm is based on instance segmentation for lane keeping on two-lane roads. Collision avoidance is performed naturally by changing lanes when obstacles are detected during driving. At this point, the curvature obtained from the bounding box of the obstacle is reflected in the driving path. In short, instance segmentation results from low-cost camera images were utilized for lane keeping, lane changing, collision avoidance, and even reused for path planning to improve safety. To validate the proposed method, the experiments are conducted in simulated and real environments and the results are presented.