Unstructured roads usually have no clear boundaries and lane lines, and the environment is more complex. The traditional segmentation methods based on road texture and color features cannot meet the requirements of real-time performance and accuracy. For unstructured road scenes, a lightweight semantic segmentation model based on improved BiSeNet was proposed, which adopted the lightweight trunk extraction network and introduced the depthwise separable convolution to optimize the speed control. The channel attention was introduced in the final feature fusion stage to adaptively select important features, suppress redundant information, and improve the accuracy of unstructured road segmentation. The number of parameters of the improved model is only 1.11×10⁶, the detection speed is increased by 18.83%, and the F1-score reaches 96.74%. Compared with other mainstream semantic segmentation models, the proposed algorithm has the advantages of small parameters, high speed and high accuracy, which can provide a reference for the safe operation of unmanned vehicles in unstructured road scenarios.