<p>Recognition of unstructured roads in agricultural fields holds significant practical value for the autonomous driving of agricultural machinery. In the field of road detection, most current research focuses on structured roads, while relatively little attention has been given to the detection of unstructured roads. Detection and navigation on unstructured farm road, a common form of unstructured roads, remain significant challenges. This paper proposes a new model that solves the problem of perceiving unstructured farm roads and adapts to the computational constraints of agricultural machinery platforms. The new model introduces an efficient attention mechanism to enhance the robustness and segmentation performance of the model, while also adopting an improved pointwise and depthwise convolution structure to reduce the model’s computational complexity and parameter count. Compared to the baseline, our model demonstrates stronger robustness under varying lighting conditions, achieving an efficiency increase to 598% of the original, along with higher segmentation accuracy.</p>