Deep LearningDriven Navigation for Autonomous Golf Carts: Implementing YOLOP for Roadside Tracking

This paper presents a vision-based navigation system for an autonomous golf cart, utilizing a custom drive-by-wire system and a single webcam strategically mounted on the left-hand side for real-time roadside tracking. Employing the You Only Look Once for Panoptic Driving Perception (YOLOP) model, optimized with TensorRT and trained on the Berkeley Deep Drive 100K dataset, the system effectively processes images to delineate drivable paths under various lighting conditions. Experiments conducted on both straight and curved paths at King Mongkut’s University of Technology Thonburi demonstrated the system’s capability to accurately track and adjust to lane deviations using a Proportional-Derivative control mechanism, ensuring precise vehicle steering and robust performance. The findings underscore the potential of integrating advanced image processing and deep learning models to enhance autonomous navigation in urban settings, suggesting avenues for further improvements in multi-sensor fusion and algorithm optimization.