Abstract: Due to the numerous features and complex structure of unstructured road, classical algorithms such as image segmentation and road model cannot meet the accuracy and real-time requirements of unstructured road identification in practical application. The above difficulties can be effectively solved by the semantic segmentation algorithm based on deep learning. The lightweight feature extraction network is adopted to improve the excessive discrete computation in the feature extraction network, optimize the control of the number of parameters and speed, and greatly reduce the redundancy of Deeplabv3+ network. According to the spatial and channel distribution characteristics of unstructured roads in the image, attention mechanism is introduced to process the high-level and low-level feature images to improve the sensitivity and accuracy of the network to feature extraction. IDD (India Driving Dataset) unstructured road data sets with sufficient and representative sample categories and strong similarity to China′s road conditions were selected to screen and preprocess the unstructured road categories and train the network. Experimental comparison shows that compared with similar networks, the number of parameters is the least, and the accuracy and rapidity are in the forefront. Compared with before the improvement, the number of network parameters decreases by 93.1%, the recognition frame rate increases by 116.1%, and the accuracy rate increases by 10.1%.