Spatial Pooling Network For Lane Line Segmentation

Abstract

In this paper, a novel method named spatial pooling is proposed for visual lane line segmentation. Occlusion is one of the most common and difficult scenes among various and complex road environments. In most existing references, the structure of lane line is still not considered for the occlusion problem. To better solve this case and enhance the long-distance information fusion for lane lines, a spatial pooling module is proposed firstly by using a combination of horizontal and vertical pooling to achieve pooling in the direction across the lane line. This method can enhance the information fusion ability in oblique directions and improve the lane line detection performance. Then, for further study, we propose an end-to-end network for lane line segmentation based on the spatial pooling module. Finally, the experimental results on tuSimple lane line dataset verify the effectiveness of our method.