Implementation of semantic segmentation for road and lane detection on an autonomous ground vehicle with LIDAR

Abstract

While current implementations of LIDAR-based autonomous driving systems are capable of road following and obstacle avoidance, they are still unable to detect road lane markings, which is required for lane keeping during autonomous driving sequences. In this paper, we present an implementation of semantic image segmentation to enhance a LIDAR-based autonomous ground vehicle for road and lane marking detection, in addition to object perception and classification. To achieve this, we installed and calibrated a low-cost monocular camera onto a LIDAR-fitted Formula-SAE Electric car as our test bench. Tests were performed first on video recordings of local roads to verify the feasibility of semantic segmentation, and then on the Formula-SAE car with LIDAR readings. Results from semantic segmentation confirmed that the road areas in each video frame were properly segmented, and that road edges and lane markers can be classified. By combining this information with LIDAR measurements for road edges and obstacles, distance measurements for each segmented object can be obtained, thereby allowing the vehicle to be programmed to drive autonomously within the road lanes and away from road edges.