半导体光电, 2020, 41 (4): 548, 网络出版: 2020-08-18
融合特征法与直接法的RGBD稠密视觉SLAM算法
RGBD Dense Visual SLAM Algorithm Combining Feature Method and Direct Method
同步定位与地图创建 直接法 特征法 建图 闭环 simultaneous localization and mapping direct method featurebased method map building loop closing
摘要
为了保持直接法的快速性与特征法的高精度和闭环能力,提出了一种融合直接法与特征法的RGBD同时定位与地图创建(SLAM)算法。该算法主要包含3个并行线程:跟踪线程、局部建图线程和闭环线程。在跟踪线程中对非关键帧进行跟踪,通过最小化光度图像误差来进行相机的初始位姿估计以及像素点的对应关系计算,利用最小化局部地图点重投影误差进一步优化相机位姿,实现快速准确的跟踪与定位;在局部建图线程中对关键帧进行提取并匹配ORB特征,执行局部BA(光束平差法),对局部关键帧位姿和局部地图点的位置进行优化,提高SLAM的局部一致性;在闭环线程中执行对关键帧的闭环检测和优化,从而保证SLAM全局一致性。另外,根据RGBD图像和相机位姿信息,通过基于Octomap的建图框架,构建完整准确的3D稠密环境地图。在TUM数据集下的实验表明,所提出的方法可以得到与基于特征法相当的精度,且所需时间更少。
Abstract
In order to maintain the fast performance of the direct method and the high precision and loop closure capability of the featurebased method, a RGBD simultaneous localization and mapping (SLAM) algorithm combining the direct method and the featurebased method is proposed. The proposed algorithm is composed of three parallel threads: tracking thread, local mapping thread and loop closing thread. In the tracking thread, the nonkey frames are tracked, the initial pose estimation and the corresponding relationship calculation of pixel points are carried out by minimizing the photometric image errors, and the camera pose is further optimized by minimizing reprojection errors of the local map points to achieve fast and accurate tracking and positioning. In the local mapping thread,the ORB features are extracted and matched features in the key frames, and the local BA (Bundle Adjustment method) is performed to optimize the position and posture of local key frames and the location of local map points, so as to improve the local consistency of SLAM. In the loop closing thread, the loop detection and the loop optimization for key frames are executed, to enhance the global consistency of SLAM. In addition, according to the RGBD image and camera pose information, a complete and accurate 3D dense environment map is constructed through Octomapbased mapping framework. Experiments on TUM datasets show that the proposed method achieves the same accuracy as featurebased method with less time.
胡章芳, 张杰, 程亮. 融合特征法与直接法的RGBD稠密视觉SLAM算法[J]. 半导体光电, 2020, 41(4): 548. HU Zhangfang, ZHANG Jie, CHENG Liang. RGBD Dense Visual SLAM Algorithm Combining Feature Method and Direct Method[J]. Semiconductor Optoelectronics, 2020, 41(4): 548.