Towards Bayesian real-time optical flow

摘要

Optical flow is a pre-requisite for computing motion detection, time to collision, structure, focus of expansion as well as object segmentation. Unfortunately, most optical flow techniques do not provide accurate and dense measures that are useful for these types of computations. In addition, most techniques are also computationally slow. Albeit, one method proposed by Camus claims and is able to perform optical flow computations in real-time capitalizing on redundancies in the computation and spatial-temporal sampling trade-offs. It is a simple technique based on simulating various motions and computing the sum–difference of patches. The shortcoming of the Camus algorithm is that the produced field is not accurate and is arbitrary for aperture and blank wall situations. However, we show that the intermediate results from the Camus approach can be used as the factored samples for the likelihood probabilities that can be used in a Bayesian spatial and temporal propagation framework. The maximization/minimization of the likelihood is not able to differentiate arbitrary from zero flow situations. It is interesting to note that the shape of the likelihood pdf clearly identifies aperture and blank wall cases. A simple diffusion of the means of reliable flow vectors produces results that are suitable for motion detection but are inaccurate for flow determination. Similar past efforts (e.g. Singh) for flow propagation produced comparable results with a computationally more complicated propagation process. It is argued that a logic is required that takes the variances into consideration in addition to the means is required for the propagation process: first propagating spatial (to address aperture and blank wall problems) and subsequently temporal information in order to maximize the number of unimodal small variance nodes.