基于BA优化和KL散度的RGB-D SLAM系统

RGB-D SLAM System Based on BA Optimization and KL Divergence

针对深度相机采集深度图像的噪声对位姿估计精度的影响,以及误差随时间累积的严重问题,设计了一种改进的基于RGB-D相机的视觉SLAM系统.首先,建立重投影误差模型,通过最小化重投影误差,对位姿和特征点进行非线性优化.此外,提出了一种闭环检测的算法,建立字典模型,用频率-逆文档频率计算权重,用Kullback-Leibler散度计算相似度,并使用相对相似度机制检测闭环,减少了累积误差.使用15个公开的图像序列对算法进行评价,同3个流行的RGB-D SLAM系统对比,精度平均最高提高119. 07%,最低提高4. 24%.实验结果证明,提出的方法比目前流行的RGB-D SLAM系统具有更好的精度.

To reduce the impact of noise in depth image acquisition on the accuracy of pose estimation and to solve the serious problem of cumulative error over time,an improved RGB-D SLAM system was designed.Firstly,the re-projection error model was established to nonlinearly optimize the poses and features by minimizing the re-projection error.In addition,a closed-loop detection algorithm was proposed.The dictionary model was established and the frequency-inverse document frequency(TF-IDF)was used to calculate the weight.Kullback-Leibler divergence was used to calculate the similarity,and a relative similarity mechanism was used for the closed-loop detection.The cumulative error was decreased.The algorithm was evaluated using 15 public image sequences.Compared with three popular RGB-D SLAM systems,the maximum increase of accuracy is averagely 119.07%and the minimum increase is 4.24%.Experimental results show that the proposed method has better accuracy than the current popular RGB-D SLAM systems.