An Adaptive Visual Odometry Based on RGB-D Sensors

Visual-based odometry is a robust technique utilized for mobile robot localization. This paper presents an adaptive RGB-D visual odometry based on the fusion of the direct methods and the feature-based methods. First, scene detection is performed to detect texture features. Then, based on the results of the scene feature evaluation function, the fusion algorithm can select the most appropriate visual odometry algorithm for motion estimation. In this paper, the details of the algorithm are discussed. In particular, the implementation of the scene feature evaluation function is described in detail. We evaluate our system on publicly benchmark to show our system has higher accuracy and better robustness.