Semantic-Based Road Segmentation for High-Definition Map Construction

Abstract

The development of autonomous driving technology proposes higher requirements of the fidelity of the high-definition maps (HD maps). The construction of HD map based on orthophotos generated from panoramic images is the state-of-the-art approach. However, in this process, the dynamic obstacles and shadows on the road captured by the panoramic camera has significant impact on reducing the map quality. Moreover, the GNSS signal may inevitably unavailable or have jitter error, leading to the unsatisfactory of the orthophoto mosaic. Therefore, an approach is proposed to tackle these problems of HD map construction. The semantic segmentation of the panoramic images is firstly implemented to extract the dynamic obstacles such as vehicles and the road segments. Then the shadows on the road segments are removed through GAN networks to generate clean orthophotos. Afterwards, the clean orthophotos are used for feature extraction and image registration based on the road segmentation to provide finer pose estimations. Finally, the GNSS data, odometer data, and estimated poses are combined to optimize the vehicle pose for orthophoto mosaic. The experimental results illustrate that the proposed approach can improve the HD map construction accuracy under the congested environment.