Geometry Embedded Multi-Task Network for Lane Detection and Classification

Abstract

Lane detection and classification are key and fundamental problems in the field of autonomous driving. Deep learning based multi-task network is commonly built with convolutional neural networks (CNNs) to handle these problems. Although existing methods have demonstrated the capability to capture the pixel-level representation and global guidance, the local geometric relationship of lanes are not taken into account to deal with the lanes with weak appearance coherence. This paper proposes geometry embedded multi-task network (GEMNet) with three branch layers, including bounding box regression, lane mask and lane classification. Convolutional operations are applied to feature maps generated by the top hidden layer of bounding box regression branch to explore the semantic context via geometric relationship. Moreover, it is able to facilitate the weight optimization in the shared layers and other branch layers simultaneously. GEMNet is evaluated on Cordova1 and Washington1 datasets, and Experimental results show that the proposed method outperforms the state-of-the art approaches. F1-Score can reach 0.906 and 0.881 with respect to Cordova1 and Washington1 datasets, respectively.