Real-Time Stereo-Vision Semantic Navigation for Apple Orchards Using YOLOv5

Abstract

Autonomous navigation in orchards is highly challenging due to dense tree canopies, variable lighting, grass, uneven terrain and frequent GNSS signal blockage. This paper presents a complete, purely stereo-vision navigation system that relies exclusively on data from a single forward-facing Intel RealSense RGB-D camera. The proposed learning-based semantic algorithm uses a YOLOv5 segmentation model to perform real-time pixel-level segmentation of the ground (road) between rows. The segmentation output is processed through a distance transform and triangle fitting to extract both heading angle and lateral offset. A hybrid steering controller with mode switching generates smooth and robust control commands while end-of-row detection is performed directly from the segmentation mask which enables fully closed-loop operation without any external positioning sensors. The system was evaluated in an apple orchard under realistic natural lighting conditions. Experimental results show a mean absolute lateral error of 0.1651 m and a root mean square error of 0.2056 m relative to the row centerline. The results indicate that accurate and stable row following can be achieved using only low-cost RGB-D data, making the proposed system a feasible navigation solution for orchard robotics.