Multi-session Joint Optimization for Accurate Mapping

Accurate maps play a vital role in the localization of mobile robots and autonomous vehicles. However, the traditional methods of creating precise maps of scenes involve special probe vehicles equipped with expensive devices, making the generalization and commercialization of accurate mapping difficult. This paper proposes a novel method to generate accurate sparse landmarks maps of the environment using a general commercial stereo camera in a multi-session pattern. It consists of a visual SLAM frontend for obtaining individual session maps and a backend for performing overlap detection and multi-session joint optimization. Experiments on the KITTI datasets validate the proposed method. It demonstrates that the proposed method can generate an accurate map only using a low-precision and low-cost stereo camera. Its accuracy is even on par with that of 3D LiDAR-based LOAM.