Trajectory-Based Stereo Visual Odometry With Statistical Outlier Rejection

We present a stereo visual odometry algorithm with trajectorical information accumulated over time and consistency among multiple trajectories of different motions. The objective function considers transfer error of all previously observed points to reduce drifting, and can be efficiently approximated and optimized within a computational bound. Different from traditional residual-based consistency measurement, we exploit the linear system in non-linear optimization to evaluate the influence of each point for outlier rejection. Both the drifting and irruptive error are reduced by combining trajectorical information of multiple motions. Experiments with real world dataset show that our method could handle difficult scenes with large portion of outliers without expensive computation.