Abstract: Academic semantic translation using computers In the autonomous operation task of smart agricultural machinery, in addition to autonomous navigation tasks, it is also necessary to pay attention to the problem of field road defect detection during autonomous operation of agricultural machinery. In the task of detecting potholes in the field, we use the publicly available conventional road pothole data on Kaggle for learning and training, and then transfer learning from the self-built farmland pothole dataset in the laboratory environment to simulate the detection of potholes in the field road. In terms of model design, we adopt the lightweight version of YOLOv5s for network design, and introduce the CBAM attention mechanism to improve network accuracy. Finally, using Jetson Nano as a carrier, we apply TensorRT to simulate the detection of potholes in the field road. The model obtained from the pothole detection task is deployed on the edge, significantly improving the running speed while appropriately reducing the accuracy, meeting the requirements of real-time detection on the edge, and realizing the edge deployment of field road defect detection technology in autonomous agricultural machinery operation tasks.

Key words: Autonomous farm machinery operations, defect detection, deep learning, lightweight model design, model edge deployment