Obstacle Detection Method for Southern Farmlands Based on Improved YOLO v8 Model

Abstract:

Obstacle detection is one of the key technologies for enabling autonomous operation of unmanned agricultural machinery. To address obstacle detection in the complex environments of unstructured farmland in southern China, a dataset of obstacles in such farmlands was collected and annotated，and data diversity was enhanced through data augmentation techniques. Based on the YOLO v8 model, the Shuffle Attention mechanism was introduced into the C2f module, proposing an improved C2f?ATT module to enhance feature representation. Simultaneously, a convolutional neural network building block was used to replace some conventional modules, forming the YO?STNet model. The enhanced intersection over union (EIoU) loss function was adopted to replace the original loss function, balancing the decreased convergence speed caused by increased model complexity. Experimental results showed that the proposed model exhibited significant advantages in detection accuracy, particularly in small object detection and blurred object recognition, while network convergence speed was also markedly accelerated. Compared with the original YOLO v8 model, the improved model achieved a 3.67 percentage points increase in average detection accuracy. Comparative results with models such as YOLO v5, YOLO v7, YOLO v10, Faster R?CNN, and SSD demonstrated that the proposed model offered superior performance and higher precision. The findings can provide important technical reference and practical support for autonomous obstacle avoidance of unmanned agricultural machinery in complex farmland environments.