Abstract: In order to meet the adaptability of autonomous vehicles to the cross-country environment and to improve the understanding ability of autonomous vehicles to the environment, higher requirements for environmental awareness system must be put forward.The most critical point in environmental awareness system is lane extraction or road extraction.However, the cross-country environment is more complicated in comparison with the structured road in the urban environment.The main reason lies in high complexity of the cross-country environment, and the extraction algorithms are different for different scene.A variety of cross-country scenes were studied, and a road segmentation method that adaptable to different scene was proposed.Firstly, a large number of data were collected for the cross-country environment, and the corresponding datasets were established.Secondly, these scenes were arranged to be identified by using deep learning method.And then a semantic segmentation algorithm was applied to segment the roads under different scenes.Finally, the whole algorithm modules were unified to obtain test results.