融合点、面特征的RGB-D视觉里程计

RGB-D Visual Odometry Fused Features of Planes and Points

基于ICP的视觉里程计算法存在计算量大、收敛缓慢的问题,并且室内环境存在大量结构化平面。为此,提出了一种融合点特征和面特征的RGB-D视觉里程计算法。首先利用改进后的ORB算法提取环境中的角点信息,然后利用层次聚类和主成分分析进行平面拟合,提取出环境中的平面特征,接下来根据特征匹配对,使用融合了点面特征的特征匹配算法解算出相机的位姿估计,最后再利用ICP算法精确计算相机的位姿变化。方法再TUM数据集中表现出了良好的性能,效果显著优于传统的ElastcFusion算法,重建结果的表面信息更为丰富准确,累积误差的精度提升,满足实时重建的要求,具有一定的实用价值。

The visual odometry based on ICP has the problems of large amount of calculation and slow convergence.There are a large number of structured planes in the indoor environment.Therefore,an RGB-D visual odometry integrating point features and surface features is proposed.Firstly,the corner information in the environment is extracted by the improved ORB algorithm,and the plane features in the environment are extracted by plane fitting using hierarchical clustering and principal component analysis.Then,according to the feature matching,the pose estimation of the camera is calculated by using the feature matching algorithm integrating point and surface features.Finally,the pose of the camera is accurately calculated by ICP algorithm.The method shows good performance in TUM data set,and the effect is significantly better than the traditional ElasticFusion algorithm.The surface information of the reconstruction results is richer and more accurate,the accuracy of cumulative error is improved,which meets the requirements of real-time reconstruction,and has certain practical value.