To achieve large-scale outdoor real-time localization, a semidirect visual odometry method for a monocular camera is proposed. In this method, the performance of the features from the accelerated segment test (FAST) algorithm is improved to detect corners with good tracking and distribution properties. A multiscale Lucas-Kanade (LK) approach is presented to build the correspondence between features and map points robustly and efficiently. Thereafter, a semidirect visual odometry system is put forward to achieve long-distance and large-scale localization for the ground vehicle. The main contribution of this method is that it integrates the robustness and efficiency advantages of feature-based methods and the accuracy virtue of direct methods into one visual odometry system. As a result, the proposed method can be applied to localize long-distance and large-scale outdoor scenes, without loop closing and global BA. Several experiments illustrate the performances of the improved FAST algorithm, the multiscale LK approach, and our semidirect visual odometry system. The experimental results demonstrate that our semidirect visual odometry system can be operated on the KITTI benchmark accurately and efficiently. The average computational time for a single frame is below 60 ms on a general notebook computer with a CPU.
Full Text Finder