Enhanced obstacle detection for intelligent agricultural machinery via transfer learning and improved YOLO11

Reliable obstacle detection is critical for the safe operation of intelligent agricultural machinery in unstructured farmland environments. However, achieving robust detection in agricultural settings remains challenging due to limited annotated data and significant domain shifts between generic visual datasets and agricultural scenes, which constrain the effectiveness of existing object detection models. To address these challenges, this study proposes an enhanced YOLO11 framework with task-aligned transfer learning for detecting obstacles in farmland. The proposed approach consists of three core components: 1) architectural refinements, including the integration of CBAM attention modules and an improved SPPF structure to enhance multi-scale feature representation; 2) task-aligned pretraining on a curated COCO subset comprising 70 552 images containing task-relevant object categories; and 3) a staged fine-tuning strategy that combines backbone freezing with subsequent end-to-end optimization on a farmland obstacle dataset. Experimental results demonstrate that the proposed method achieves a mAP@0.5 of 0.934 on the test set, improving mAP@0.5 by 6.4 percentage points over YOLO11-s. Furthermore, after TensorRT optimization, the model reaches 82.6 FPS on the Jetson AGX Orin platform while maintaining a mAP@0.5 of 0.927, confirming its suitability for real-time deployment. These findings indicate that task-aligned transfer learning, combined with targeted architectural enhancements, effectively mitigates data scarcity and domain shift in agricultural obstacle detection.