Inclined Obstacle Recognition and Ranging Method in Farmland Based on Improved YOLOv8

Abstract

Unmanned tractors under ploughing conditions suffer from body tilting, violent shaking and limited hardware resources, which can reduce the detection accuracy of unmanned tractors for field obstacles. We optimize the YOLOv8 model in three aspects: improving the accuracy of detecting tilted obstacles, computational reduction, and adding a visual ranging mechanism. By introducing Funnel ReLU, a self-constructed inclined obstacle dataset, and embedding an SE attention mechanism, these three methods improve detection accuracy. By using MobileNetv2 and Bi FPN, computational reduction, and adding camera ranging instead of LIDAR ranging, the hardware cost is reduced. After completing the model improvement, comparative tests and real-vehicle validation are carried out, and the validation results show that the average detection accuracy of the improved model reaches 98.84% of the mAP value, which is 2.34% higher than that of the original model. The computation amount of the same image is reduced from 2.35 billion floating-point computations to 1.28 billion, which is 45.53% less than the model computation amount. The monitoring frame rate during the movement of the test vehicle reaches 67 FPS, and the model meets the performance requirements of unmanned tractors under normal operating conditions.