Continuous-Time Stereo Visual Odometry Based on Dynamics Model.

We propose a dynamics model to represent the camera trajectory as a continuous function of time and forces. Equipped with such a representation, we convert the classical visual odometry problem to analyzing the forces applied to the camera. In contrast to the classical discrete-time estimation strategy, the continuous nature of the camera motion is inherently revealed in the framework, and the camera motion can be simply modeled with only few parameters within time intervals. The dynamics model guarantees the continuous velocity, and hence assures a smooth trajectory, which is robust against noise and avoiding the pose vibration. Evaluations on real-world benchmark datasets show that our method outperforms other continuous-time methods.