Stereo odometry based on careful feature selection and tracking

In this paper we present a novel algorithm for fast and robust stereo visual odometry based on feature selection and tracking (SOFT). The reduction of drift is based on careful selection of a subset of stable features and their tracking through the frames. Rotation and translation between two consecutive poses are estimated separately. The five point method is used for rotation estimation, whereas the three point method is used for estimating translation. Experimental results show that the proposed algorithm has an average pose error of 1.03% with processing speed above 10 Hz. According to publicly available KITTI leaderboard, SOFT outperforms all other validated methods. We also present a modified IMU-aided version of the algorithm, fast and suitable for embedded systems. This algorithm employs an IMU for outlier rejection and Kalman filter for rotation refinement. Experiments show that the IMU based system runs at 20 Hz on an ODROID U3 ARM-based embedded computer without any hardware acceleration. Integration of all components is described and experimental results are presented.