基于混合特征的机器人定位与地图创建

Robot localization and map building based on hybrid features

针对原始ORB-SLAM算法仅依靠点特征进行位姿估计,在点特征不足场景下会导致算法精度和鲁棒性降低的问题,对原有算法进行改进。首先,构建二进制环境词典用于提高算法加载词典速度;其次,在剔除特征误匹配时,提出一种改进随机采样一致算法(RANSAC)以提高剔除效率;然后,构建混合特征模型估计位姿;最后,根据图优化模型,提出基于最大共视权重帧的全局Bundle Adjustment算法用于位姿优化并实现稠密和稀疏地图创建。结果表明,混合特征模型能有效减少算法误差。采用二进制环境词典加载时间由14.945 s减小至0.420 s;改进RANSAC算法执行时间由0.803 s减小到0.078 s;采用TUM数据集对算法进行检验,均方根误差分别下降2.27、4.87 cm。根据数据集上的实验结果证明了所提算法有效性。

Aiming at the problem that the original ORBSLAM algorithm only uses point features to estimate pose, which leads to the degradation of the algorithm accuracy and robustness in the scene with insufficient point features. In order to improve the efficiency and accuracy of the algorithm, the original algorithm is improved. Firstly, this paper builds a binary environment dictionary, which can increase the dictionary loading speed of the algorithm. Secondly, an improved RANSAC algorithm is proposed to increase the elimination efficiency in eliminating feature mismatching. Then, a hybrid feature model is established to extract the point features and line features in the image to estimate pose. Finally, according to the pose graph optimization model, a global Bundle Adjustment algorithm based on the maximum common view weight frames is proposed and used in pose optimization, and the establishing of the dense and sparse maps is realized. The experiment results show that the hybrid feature model can effectively reduce the algorithm error. With the binary environment dictionary, the loading time of the dictionary is decreased from 14.945 seconds to 0.420 seconds. The execution time of the improved RANSAC algorithm is reduced from 0.803 seconds to 0.078 seconds. Using TUM dataset to verify the algorithm, the root mean square error of the camera trajectory is decreased by 2.27 cm and 4.87 cm, respectively. The experiment results on the dataset prove the effectiveness of the proposed algorithm.