Abstract:

Simultaneous localization and mapping （SLAM） is an essential part of autonomous vehicles. Its existing algorithms are mainly based on lidar or visual-inertial odometer， which do not make full use of the respective advantages of multi-mode sensors， leading to insufficient robustness to featureless scenes. In view of these， a multi-sensor tightly coupled SLAM system using lidar， camera and inertial measurement unit is proposed in this paper. Firstly， the system improves the schemes of the feature extraction of lidar point cloud and plane fitting and enhances the efficiency and accuracy of the depth information optimization of visual feature points by using lidar point cloud. Secondly， the tightly coupled state estimation framework proposed directly adds lidar odometer constraints onto visual inertial system， so enhancing the system stability and accuracy without increasing the complexity of algorithm. Finally， the coarse-to-fine visual-lidar coupled loop framework further reduces the long-term cumulative drift of the system. The results of massive tests for validation on the open-source dataset KITTI show that compared with other commonly used algorithms， the proposed algorithm achieves higher accuracy and environmental adaptability. In addition， the real vehicle test on the self-built autonomous vehicle test platform also demonstrates its adaptability to the long-time and large scene environment.

Key words: autonomous vehicles, state estimation, SLAM, multi-sensor-fusion, closed-loop detection