融合K-Means和多尺度自适应的复杂场景视觉SLAM方法

A Visual SLAM Method for Complex Scenes Combining K-Means and Multi Scale Adaptive

针对传统视觉SLAM前端误匹配使定位精度低和追踪失败导致系统鲁棒性下降问题,提出了一种融合K-Means聚类和金字塔多尺度自适应的解决方法,并应用于不同复杂场景环境中。首先,算法在ORB-SLAM2结构上改进,将K-Means聚类算法融合到匹配算法里,一方面基于聚类的思想减少具有更多特征点的关键帧处理事件,解决了误匹配和匹配精度低的问题。另一方面,系统跟丢情况下,在重定位阶段融合多尺度自适应方法进行重定位,提高系统鲁棒性。在公开数据集上的实验结果表明,与其他匹配方法对比,提出的改进K-Means匹配剔除算法在最短的时间内保留原始算法2倍以上的正确匹配对,多尺度自适应重定位相比ORB-SLAM2在复杂场景下重定位效果优异;整体性能方面,改进后的系统相比原始算法精度平均提升了30%,系统耗时平均降低1 s。改进后的轨迹更为接近实际运动轨迹,展示了方法的有效性。

Aiming at the problem of low positioning accuracy and poor system robustness due to tracking failure caused by traditional visual SLAM front-end mismatch, a solution combining K-Means clustering and pyramid multi-scale adaptive is proposed and applied to different complex scene environments. First of all, the algorithm is improved on the ORB-SLAM2 structure, and K-Means clustering algorithm is integrated into the matching algorithm. On the one hand, based on the idea of clustering, key frame processing events with more feature points are reduced, which solves the problem of wrong matching and low matching accuracy. On the other hand, when the system is lost, the multi-scale adaptive method is fused in the relocation stage to improve the system robustness. The experimental results on the public dataset show that, compared with other matching methods, the proposed improved K-Means matching elimination algorithm retains more than twice the correct matching pairs of the original algorithm in the shortest time, and the multi-scale adaptive relocation is superior to ORB-SLAM2 in the relocation of complex scenes;in terms of overall performance, the accuracy of the improved system is increased by 30% on average compared with the original algorithm, and the system time consumption is reduced by 1 s on average. The improved trajectory is closer to the actual trajectory, which shows the effectiveness of the method.