VEst: An Efficient Solution to the Camera Velocity Estimation from Minimal Feature Points
Yujie Zhang, Kaveh Fathian, Nicholas Gans · 2020
This paper presents a new method to determine camera velocity from image data. Velocity estimation has been underexplored, compared to estimating rotation and translation. Existing methods suffer from various drawbacks in terms of necessary conditions on the 3D structure of the scene, the types of camera velocities they can solve for, convergence of estimates, or necessary partial knowledge of the scene or velocity. Our approach uses optical flow vectors and delivers estimates of angular and linear velocities. Rigid body velocity equations are used to derive a polynomial system, which is then solved for the desired velocity terms as well as the current depths of feature points that root the optical flow vectors. Simulations are conducted to validate the correctness and accuracy of the proposed algorithm. The real-world KITTI dataset is used further to demonstrate the performance in practical problem.