Development of Sensor Fusion-based Obstacle Detection and Collision Avoidance Technology for Autonomous Tractor

Abstract

For the practical implementation of autonomous agricultural machinery, reliable obstacle recognition and collision prevention systems are essential in agricultural environments. This study aimed to develop a collision prevention system for autonomous tractors capable of real-time obstacle recognition and responsive collision avoidance using video sensor fusion technology. Emphasizing human safety as the top priority, the system targeted obstacle recognition specifically for humans and operated within designed risk and warning zone ranges aligned with the tractor's direction. The developed sensor fusion system integrated the robust object classification capabilities of an RGB camera with a lidar sensor for precise distance measurement, achieving synchronization through camera-lidar calibration. The recognition algorithm, based on the YOLO model applied to RGB images, identified humans, while the lidar data projected onto the image measured the relative distance from the tractor. The collision prevention system, relying on the relative distance of recognized obstacles, halted the tractor in hazardous areas, resuming operation once the obstacle cleared. Validation in static and dynamic situations on flat farmland demonstrated a recognition rate exceeding 99% in collision risk zones, a 24 cm RMSE for distance measurement, and a success rate of over 98% in collision prevention and response.