Monocular SLAM for large scale scenes

Abstract

Monocular simultaneous localization and mapping (SLAM) methods easily accumulate error along with the growing map. Especially for a large scale scene, accumulating error may be too large to prevent the close of loops. Bundle adjustment can be employed to close loops and eliminate accumulation error. However, for a large scale scene, traditional bundle adjustment may have the bottleneck in terms of memory and efficiency, which limits its applications in practice. Based on our prior off-line SfM work, in this paper, we present a novel monocular SLAM system for large scale scenes. We modify and improve the key techniques to make them appropriate for real-time SLAM. Particularly, a novel online loop closure detection method based on non-consecutive track matching is proposed, which can automatically detect and efficiently match the common features among different subsequences. In addition, a segment-based bundle adjustment is employed to close the loop efficiently, which can perform an efficient global optimization in a limited memory space to effectively eliminate accumulation error. In the experiments, we made qualitative and quantitative comparison with two state-of-the-art SLAM systems using both our own dataset and KITTI dataset, which demonstrates the effectiveness of the proposed approach.