Optimal Estimation of Optical Flow, Time-to-Contact and Depth

摘要

Extraction of the optical flow field, time-to-contact and depth from an imagesequence is a vitally important problem in many contexts. Many robotic applications, e.g., obstacle avoidance, autonomous vehicles and the like, can benefit from this information. In this report, the problem is formulated under a least squares minimization framework. Based only upon a set of spatiotemporal disparity measurements from a correlation algorithm, the algorithm in general recovers optimal estimates, in the sense of least squares, of both the instantaneous optical flow field and time-to-contact simultaneously. It does not utilize knowledge of camera translation or any internal camera parameters. Nevertheless, additional information can be recovered if prior information is incorporated into the algorithm. It is shown that if complete translational camera motion as well as some internal camera parameters are known, optimal estimates of absolute depth can directly be inferred from the disparity measurements. Closed-form solutions are obtained in all cases. The algorithm is simple, effective, and easy to implement, which enhances the possibility of hardware realization for real-time applications. Experimental results, both synthetic and real, are included to show the performance of the algorithm and to verify its robustness against random noise perturbation.