For a moving robot based on visual Simultaneous Localization and Mapping (SLAM), blurred images will degrade the accuracy of localization. Therefore, how to handle blurred images is a main problem in visual SLAM. In order to decrease the influence of blurred images on localization accuracy, this paper proposes an improved visual SLAM, which is based on Haar wavelet transform and has the ability of eliminating blurred images. Besides, a correlation-weighted pose optimization is also developed in this paper. This weighted optimization integrates the correlation between matching features as weighting coefficients into the reprojection errors. In this weighted method, pose optimization algorithm can reduce the influence of the matching features with low correlation, which are more likely to be mismatched. As a result, the accuracy of the estimated pose will be improved. The improved system optimized by our method is evaluated on the TUM RGB-D dataset and real environment. It is also compared with other optimization systems, which were based on blurred image elimination and uncertainty-weighted optimization respectively. The experimental results demonstrate that the system optimized by our method could achieve the highest accuracy and robustness in pose estimation.

对于基于视觉同步定位和映射 (SLAM) 的移动机器人，模糊的图像会降低定位的准确性。因此，如何处理模糊图像是视觉SLAM中的一个主要问题。为了减少模糊图像对定位精度的影响，本文提出了一种改进的视觉SLAM，它基于Haar小波变换，具有消除模糊图像的能力。此外，本文还开发了相关加权姿态优化。这种加权优化将匹配特征之间的相关性作为加权系数整合到重投影误差中。在这种加权方法中，位姿优化算法可以减少相关性低的匹配特征的影响，这些特征更容易失配。因此，估计姿态的准确性将得到提高。通过我们的方法优化的改进系统在 TUM RGB-D 数据集和真实环境上进行评估。还与其他优化系统进行了比较，这些优化系统分别基于模糊图像消除和不确定性加权优化。实验结果表明，通过我们的方法优化的系统可以在姿态估计中达到最高的准确性和鲁棒性。