基于双目相机的SLAM回环检测改进算法研究

Improved Algorithm of SLAM Loop Closure Detection Based on Binocular Camera

回环检测算法在视觉即时定位与地图构建(SLAM)系统框架中具有重要的作用,能够修正在前端和后端计算过程结果中产生的不可避免的累计误差,在长时间运行的情况下能够保证定位与建图的准确性。传统的视觉SLAM中主要采用基于词袋模型的图像检索方式,但是在视角差异较大的情况下难以匹配到正确的图像。针对这种情况,基于双目相机提出利用透视变换和基于迭代最近点(ICP)算法的空间信息几何检验,对候选关键帧进行视角变换和匹配,完成较大视角差异情况下的回环检测,修正累计误差。该方法适用于对算法耗时要求不甚严格的场景。实验结果表明,该方法可以提高识别回环检测候选关键帧的准确率和召回率。

The loop closure detection algorithm plays an important role in the visual simultaneous localization and mapping(SLAM) system framework.The unavoidable cumulative errors in the results of the front-end and back-end calculation processes can be corrected, and the accuracy of positioning and mapping in the case of long-term operation can be ensured.The traditional visual SLAM mainly uses the image retrieval method based on the bag-of-words model, but it is difficult to match the correct image when the perspective difference is large.To address this situation, the use of perspective transformation and geometric inspection of spatial information based on iterative closest point(ICP) algorithm based on binocular camera is proposed to transform and match the viewpoint of candidate key frames to complete the loop closure detection in the case of large viewpoint difference and correct the accumulated error.The method is suitable for scenarios where the algorithm time consumption requirement is not particularly strict.The experimental results show that this method can improve the accuracy and recall rate of identifying loop closure detection candidate key frames.