Enhancing YOLOPv2 Lane Detection to Gain Curvature Radius

Lane detection is essential for Advanced Driver Assistance Systems (ADAS) and autonomous driving, enabling vehicles to perceive lane boundaries, and accurately ensuring safe and reliable navigation. Recent research on lane detection on ADAS has more comprehensive data output which is critical for effective lane-keeping and vehicle control. This paper explores the enhancement of YOLOPv2 with Bird's Eye View (BEV) post-processing for lane detection tasks and compares its performance with PolyLaneNet. The study is conducted on a Jetson Orin NX platform, Mini PC RTX 2060, and PC with RTX 4060, comparing the FPS performance of both models. YOLOPv2, enhanced with BEV postprocessing, achieves 3.43 FPS in Orin NX, 11.22 FPS in RTX 2060, 18.34 FPS in RTX 4060 on custom video data while providing detailed lane information, including lane curvature, curvature radius, and lane deviation. In contrast, PolyLaneNet, which provides higher FPS, only offers lane curvature without additional detailed metrics. This study highlights the benefits of BEV post-processing on YOLOPv2's performance and its comprehensive data output, which is critical for effective lane-keeping and vehicle control.