Multi-stage Attention-Pooling Network for Lane Marking Detection

The problem of lane marking detection is one of the most exigent tasks involved in autonomous driving. It is not a straight task to solve as there are several barriers associated with the lane markings such as trivial appearance, irregular painting, intricate shapes and less spatial coverage which trigger the traditional convolutional neural networks (CNN) to extract numerous false positives from the background. In this paper, we present a novel segmentation model named MAP-Net, which focuses on extracting fine level lane marking features and produces richer semantic consistencies in detecting them. MAP-Net utilizes the integration of attention and multi-stage pooling to aggregate only the relevant features after analysing the image features which are extracted in a hierarchical manner. Further, the attention gates process the information extracted by the encoder and allow only the lane pertinent features to pass through them, thus reducing the false positive rate. We test our approach on two publicly available yet challenging datasets (Tusimple and VPGNet). The experimental results demonstrate that MAP-Net's performance is on-par and also exceeds certain benchmark methods and the predictions show improved semantic consistencies in detecting the lane markings.