What does ground tell us? Monocular visual odometry under planar motion constraint

Recently, visual odometry has gained more attention due to its application to autonomous vehicles and robots. In this paper, we propose visual odometry using a single camera. This configuration is economically more attractive than other popular configurations using stereo cameras. To estimate frame-to-frame motion more accurately, we thoroughly utilize prior knowledge on our operating space and motion. Our visual odometry assumes that its operating space is locally flat so that its movement undergoes planar motion. From this assumption, we can adopt the novel normalized 1-point algorithm to estimate relative pose between a pair of images. Moreover, under this assumption, we can resolve scale ambiguity, the most serious problem in monocular visual odometry. To estimate scale of motion, we introduce the other 1-point algorithm derived from planar homography of the ground plane. Two 1-point algorithms are the key to accelerate our visual odometry for its further application to real-time systems. Our experiments on synthetic data verified that the normalized 1-point algorithm was more accurate than other known method. Moreover, our experiments on real data presented effectiveness of our proposed visual odometry.