A filtering approach for computation of real-time dense optical-flow for robotic applications

Abstract

The present article presents an iterative filter approach for the computation of optical-flow. The filter is based on an update and propagation loop. The propagation stage takes the currently computed flow to predict the value at the next time iteration. The update stage takes this prediction, together with the stream of images, and corrects the optical-flow field. This leads to an incremental approach to build optical-flow. Regions of the image where no flow computation is possible are filled in by a diffusion term in the propagation step. Ground truth validation of the algorithm is provided by simulating a high-speed camera in a 3D scene. The local computations and convolution based implementation is well suited for real-time systems with high-speed and high-acuity cameras.