Traffic scene semantic segmentation using self-attention mechanism and bi-directional GRU to correlate context

Context information plays an important role in semantic segmentation of urban traffic scenes, which is one of the key tasks of the intelligent platform's (such as unmanned vehicles) perceiving environment, and has inspired a wide range of interests from researchers. This paper synthesizes three considerations: feature space correlation, information distributed in the long distance of image plane and long distance sequence information, and proposes a combination of self-attention mechanism and bi-directional gated recurrent unit (GRU) neural network to extract various contextual information on the basis of deep feature network, so as to achieve better semantic segmentation performance. In order to explore the optimal implementation, two kinds of topological connections are attempted. One is self-attention branch and bi-directional GRU branch in series, and the other is in parallel. In addition, in order to train the network better and achieve more precise segmentation results, a cascade refinement supervised method using two losses is proposed. Experiments carried out on Cityscapes, Mapillary, CamVid and KITTI semantic segmentation datasets demonstrate the outstanding performance and robust generalization ability of our method.