Abstract: Lane detection and recognition is one of the key technologies in automatic driving and vehicle safety early warning system. Convolutional neural network is a feature extraction method that can effectively characterize the semantic features of regions, providing a new idea for lane detection. Based on the theory of in-depth learning, the lane detection research was carried out, a new lane detection approach, which can adapt to complex and changeable driving environment was explored. A lane detection algorithm based on the convolution neural network theory of SegNet was proposed. Firstly, target detection was carried out, lane data set was generated by annotation. An annotation algorithm was designed to build a fine-labeled lane data set for lacking of lane data set. Subsequently, six coefficients labels, three for each lane, were established to redraw the lane. Combined with conditional random field post-processing method ,the network model was trained by the lane data set ,which was provided by self-annotated and MLND-Capstone open source project in CSDN. Then, I test the performance of the built Segnet network model in different road environments. Compared with the traditional lane detection algorithm based on threshold segmentation and curve fitting. The method based on deep learning reduced the interference of illumination changes, driving vehicles, road damage and other factors, and improved the robustness and detection accuracy of the algorithm. Key words: Lane detection; Deep learning; threshold segmentation; curve fitting; Segnet network

CICTP