Revisiting visual odometry for real-time performance

Abstract

Visual Odometry (VO) is a key component in modern driver assistance systems and robotics. Meeting the real-time requirements is mandatory for VO in such applications. Previous works have primarily focused on improving accuracy at the cost of longer runtime. In this work, we propose novel strategies for feature correspondence setup, outlier removal and robust pose optimization in the VO pipeline to achieve real-time performance of close to 30 frames-per-seconds (fps) on a dual-core 3.5 GHz CPU while maintaining high accuracy. In particular, computationally efficient strategies are introduced to obtain an initial set of good features and rapidly filter out the outliers to minimize the computational overhead in later stages. In addition, we propose a depth based weighting and saturated-residual scheme during pose optimization to increase the robustness of VO. Experimental results show that the proposed VO achieves the fastest speed among all the top-ranked OV and SLAM systems on KITTI leader-board. Specifically, the proposed VO is 47% faster than state-of-the-art ORB-SLAM2 with comparable accuracy on KITTI dataset.