Visual perception for autonomous robots is one of the key technologies to realize robot-environment interaction. Lane, as a kind of geometry information to ensure the safety of robot motion , has a wide research value, and lane detection in semi-structured scenarios brings new challenges to this area. The conventional method based on classic feature extraction process does not show robustness for non-urban environments or complex scenarios that have only degenerate geometric features. The use of deep learning methods for lane detection has become a mainstream trend. In this paper, we review the research on visual edge detection for mobile robots in semi-structured scenarios, and the application prospect and effect of the lane detection algorithm in semi-structured scenarios are investigated. First, the common edge line detection datasets are organized, and the current datasets and the focus of the research are analysed from the perspective of acquisition scenarios and annotation types. Second, different methods are classified and summarized, and the detection and data processing processes of different methods are compared. Then, the commonly used evaluation metrics of deep learning are sorted out, and the detection results of different methods in the face of different scenarios are compared and analyzed. Finally, the research direction of the edge detection method based on deep learning is prospected for the problems of visual lane detection in semi-structured scenes.