Road Segmentation based on Deep Learning with Post-Processing Probability Layer

Abstract

In order to solve the problem of road image semantics segmentation, a novel semantic segmentation method based on deep learning is proposed in this paper, and it is verified that this method can optimize the segmentation effectively comaring with traditional Full Convolution Neural Network (FCN) model. Firstly, the traditional Full Convolution Neural Network model is constructed. And then according to the principle of the post-processing probability layer method proposed in this paper, the label of all road image training sets is used to compute and transform it to form a two-dimensional array which can represent the classification probability of each pixel, and it is combined with the Full Convolution Neural Network model to be used for road image semantics segmentation. Secondly, the tensorflow neural network framework is used to simulate the above two models. Finally, the experimental results show that the CNN model with the proposed post-processing probabilistic layer is able to get better results in road semantics segmentation in KITTI data sets. The pixel accuracy is improved from 88.8% to 91.3%, and the mean pixel accuracy is increased from 82.9% to 85.7%. The mean intersection over union increased from 72.5% to 77.9%.