Rigid Multi-Motion Optical Flow Estimation
thesisposted on 06.08.2018 by Tomas Gerlich
In order to distinguish essays and pre-prints from academic theses, we have a separate category. These are often much longer text based documents than a paper.
Optical flow is a fundamental problem in computer vision. Given a pair of image frames from a video source, the objective is to estimate the movement, or translation of pixels in the second image with respect to the first image. Various scene understanding applications rely on accurate optical flow estimation which is often built into early stages of their computation pipelines. Our primary focus is on enabling traffic monitoring applications where optical flow can be used to track vehicles in traffic imagery capturing urban roads or highways. Optical flow has been studied extensively in the past, however estimating flow fields in realistic scenes containing large distance motion along with textureless regions, specular reflections or shadows, remains a challenging problem. Because optical flow is an under-constrained problem, the large number of methods often differ in at least two respects: the specification of the smoothness constraint in the solution and the associated minimization procedure to reach a near global optimum. In this dissertation, we approach the optical flow problem by leveraging prior knowledge of the types of motions expected to be observed in traffic imagery. In particular, vehicles follow rigid motion with a small inter-frame rotation component. We take advantage of projective geometry constraints, where an estimate of the epipolar geometry of each motion allows us to reduce the search space for pixel correspondences from a 2D plane to a 1D line. We present a novel optical flow algorithm and show that embedding epipolar constraints within the flow estimation procedure yields significantly more accurate optical flow estimates compared to state-of-the-art general purpose algorithms. We evaluate the algorithm on our novel benchmark dataset with precise ground truth and on imagery from real traffic cameras.