Hybrid-Residual-Based RGBD Visual Odometry

Visual odometry has greatly progressed since non-linear optimization methods were introduced for pose estimation. Furthermore, RGBD visual odometry has become a hot research topic in the robotic and computer vision field with the introduction of RGBD cameras. However, most RGBD-camera-based visual odometry methods are designed by extending monocular visual odometry methods, therein not paying much attention to the integration of the different types of information provided by RGBD images. In this paper, we propose a novel hybrid-residual-based-RGBD visual odometry, where three types of complementary information are integrated into a joint optimization model. The reprojection residuals, the photometric residuals and the depth residuals are minimized together in the non-linear optimization process, where a robust cost function and outlier filtering are employed in the iterative optimization to enhance the robustness of the iteration while simultaneously maintaining the optimality. Experiments on publicly available RGBD data sets validate the advantages of the integration of multiple types of information for RGBD visual odometry. The accuracy and robustness are greatly improved by our method.