徐 岩，安卫凤
AuthorsHTML:徐 岩，安卫凤
AuthorsListE:Xu Yan，An Weifeng
AuthorsHTMLE:Xu Yan，An Weifeng
Unit:天津大学电气自动化与信息工程学院，天津 300072
Unit_EngLish:School of Electrical and Information Engineering，Tianjin University，Tianjin 300072，China
Abstract_Chinese:同时定位与地图构建(SLAM)在智能驾驶和机器人技术中发挥着重要的作用．针对传统随机抽样一致(RANSAC)算法对噪声敏感的问题，提出了一种改进的RANSAC算法，命名为LORANSAC，简称LO^*．该算法包含内点筛选和非线性优化两部分．首先，在传统RANSAC算法估计出较好的模型后，保存在这个模型下得到的内点，在这些内点中随机选出一个子集，以进一步缩小内点的选择范围，迭代地进行模型估计．然后，对估计的模型进行捆集调整，通过最小化误差优化模型．实验使用公开的TUM RGBD数据集和KITTI数据集中的共10个序列进行评估，每个序列至少存在一个闭环，数据集涵盖小型和大型、室内和室外环境．从定性角度验证该算法删除误匹配的特征点的有效性，从定量角度验证使用该算法的定位精度．实验结果显示，与传统的RANSAC算法相比，改进的算法可以提高SLAM的定位精度．此外，实验结果与4个流行的SLAM系统对比，精度平均最高提高60.82%，最低提高12.16%．实验结果证明，该方法可以有效提高SLAM的定位精度．
Abstract_English:Simultaneous localization and mapping(SLAM)has been playing an important role in intelligent driving and robotics．To address the noise sensitivity problem of the traditional random sample consensus (RANSAC) algorithm，an improved RANSAC algorithm called locally optimized RANSAC(LORANSAC，abbreviated as LO^*)is proposed．The algorithm comprises two parts：inners screening and nonlinear optimization．First，after the traditional RANSAC algorithm estimates a good model，the inners produced by this model are saved．Then，a subset is randomly selected among these inners to further narrow the selection range of the inners and iteratively perform model estimation．Finally，the estimated model is bundled and optimized by minimizing the error．The proposed algorithm is evaluated using 10 public sequences from the TUM RGBD and KITTI datasets，with at least one closed loop for each sequence，which cover both small-and large-scale indoor and outdoor environments．Experiments were conducted to qualitatively validate the algorithm in deleting the incorrect matches of feature points and quantitatively verify the positioning accuracy of the algorithm．The experiment results show that in comparison with the traditional RANSAC algorithm，the improved algorithm can enhance the positioning accuracy of SLAM．In addition，in comparison with the average accuracy of the four well-known SLAM systems，the average accuracy of the improved algorithm is increased by a maximum of 60.82% and a minimum of 12.16%．The experiment results show that the proposed method can effectively improve the positioning accuracy of SLAM．
Keyword_Chinese:机器视觉；同时定位与地图构建；随机抽样一致
Keywords_English:machine vision；simultaneous localization and mapping；random sample consensus

PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6531