Performance Evaluation of YOLOv5 for Object Detection in Agricultural Implement Changeover

Abstract

The automation of agricultural implement changeovers is crucial for minimizing human intervention in the operation of autonomous farming systems. To ensure system safety and resilience, it is imperative to recognize implements as non-obstacle objects, thereby facilitating the seamless hitching of implements with autonomous tractors. This study presents the initial step in developing a safety function for autonomous implement changeover by assessing the performance of YOLO-based detectors, primarily in terms of precision and speed in detecting target implements and humans. These detectors are trained using transfer learning, employing four YOLOv5 variants (YOLOv5s, YOLOv5m, YOLOv5l, and YOLOv5x) and a custom dataset comprising 26,661 labeled images across nine classes of implements and eight classes of obstacles and equipment. The training results show a high average precision (AP) of the detectors, varying from 0.907 to 0.995, for detecting the implements. The mean average precision (mAP@0.5) for detecting all classes ranged from 0.955 to 0.966. Furthermore, testing involving tractor-implement alignments demonstrates the rapid detection of implements and humans by all detectors, with average inference times varying from 7.0 to 20.5 ms. These detectors consistently provide accurate predictions for target objects, with confidence scores (CS) varying from 87.6% to 90.4%. Notably, the detector trained with the medium-variant YOLOv5m is the optimal model with overall performance in terms of both detection speed and accuracy.