Detecting Motion Patterns in Dense Flow Fields: Euclidean Versus Polar Space

Abstract

AbstractThis research studies motion segmentation based on dense optical flow fields for mobile robotic applications. The optical flow is usually represented in the Euclidean space however, finding the most suitable motion space is a relevant problem because techniques for motion analysis have distinct performances. Factors like the processing-time and the quality of the segmentation provide a quantitative evaluation of the clustering process. Therefore, this paper defines a methodology that evaluates and compares the advantage of clustering dense flow fields using different feature spaces, for instance, Euclidean and Polar space. The methodology resorts to conventional clustering techniques, Expectation-Maximization and K-means, as baseline methods. The experiments conducted during this paper proved that the K-means clustering is suitable for analyzing dense flow fields.KeywordsVisual perceptionOptical flowMotion segmentation