Abstract:Planar map showing roads information are always incomplete. In order to improve the diversity of planar maps, it is necessary to accurately segment the road regions. Therefore, a method for improving convolutional neural networks is proposed to segment the planar map region. This experiment selects two road information-rich databases, namely Baidu database (Baidu) and Gaode database (Amap) , and then mark the pixel training set with the tag information. Using the Sigmoid segmentation objective function instead of the complex Softmax function, the Baidu-CNN model and the Amap-CNN model are trained respectively.And adjust the pixel probability with nonlinear mapping, so the fuzzy inference system is constructed. The nonlinearly mapped pixel point probability is input into the fuzzy inference system to determine the probability that the pixel point belongs to the road region, and the road segmentation result is obtained. The results show that the planar map road segmentation model obtained by the proposed algorithm has better segmentation effect. The accuracy rate can reach 94.49%, and the road segmentation speed of a single planar map can reach 5 seconds.