Stereo Vision Based Obstacle Avoidance Local-Planner for Autonomous Golf Cart

Perception and planning are key factors in deter-mining the safety of an autonomous vehicle. This paper proposes an approach for obstacle avoidance local planner by utilizing a stereo camera. Obstacles are detected by acquiring bounding boxes from an object detection architecture, i.e. the Faster R-CNN model. The boxes are then matched with depth data from the stereo camera in order to generate object points. In addition, a conformal state lattice is applied for the local planner to generate paths based on the object points. The proposed perception and planning system is implemented experimentally in an autonomous golf cart. Experimental results show that the system is able to detect and avoid static obstacles satisfactorily. Moreover, the autonomous golf cart is able to reach the desired final point with an error of less than 1 m and avoids an object with an average distance of 2 m.