We propose Stereo Direct Sparse Odometry (Stereo DSO) as a novel method forhighly accurate real-time visual odometry estimation of large-scaleenvironments from stereo cameras. It jointly optimizes for all the modelparameters within the active window, including the intrinsic/extrinsic cameraparameters of all keyframes and the depth values of all selected pixels. Inparticular, we propose a novel approach to integrate constraints from staticstereo into the bundle adjustment pipeline of temporal multi-view stereo.Real-time optimization is realized by sampling pixels uniformly from imageregions with sufficient intensity gradient. Fixed-baseline stereo resolvesscale drift. It also reduces the sensitivities to large optical flow and torolling shutter effect which are known shortcomings of direct image alignmentmethods. Quantitative evaluation demonstrates that the proposed Stereo DSOoutperforms existing state-of-the-art visual odometry methods both in terms oftracking accuracy and robustness. Moreover, our method delivers a more precisemetric 3D reconstruction than previous dense/semi-dense direct approaches whileproviding a higher reconstruction density than feature-based methods.
展开▼
