DYNAMIC TEXTURE RECOGNITION USING OPTICAL FLOW FEATURES AND TEMPORAL PERIODICITY.pdfVIP

DYNAMIC TEXTURE RECOGNITION USING OPTICAL FLOW FEATURES AND TEMPORAL PERIODICITY Sa?ndor Fazekas and Dmitry Chetverikov Computer and Automation Research Institute, Budapest, Kende u.13-17, H-1111 Hungary ABSTRACT We address the problem of dynamic texture (DT) classifi- cation using optical flow features. Optical flow based ap- proaches dominate among the currently available DT recogni- tion methods. We introduce rotation- and scale-invariant DT features based on local image distortions computed via optical flow. Then we describe an SVD-based method for measuring the degree of temporal periodicity of a dynamic texture. Fi- nally, we present the results of a DT classification study that compares the performances of different flow features for nor- mal and complete optical flows. 1. INTRODUCTION Following the terminology of the pioneering paper [1], three visual motion pattern categories are usually distinguished, na- mely, motion events, activities, and dynamic (temporal) tex- tures. Individual motion events, like a single opening of a door, possess no temporal or spatial periodicity; they are be- yond the scope of this study which is devoted to periodic- ity analysis. Activities of moving non-rigid objects, like dig- ging, are periodic visual motion patterns localised in space and time. Finally, dynamic textures are spatially repetitive, time-varying motion sequences exhibiting certain temporal stationarity. This assumes temporal repetitiveness, which is a basis of most dynamic texture models (e.g., [2]) and meth- ods for DT recognition [3]. A generic property of DTs is their intrinsic dynamics, that is, the motion that cannot be compensated by any global spatial transformation. Translational motion of an otherwise static texture does not generate a real (strong) dynamic tex- ture. At most, it can be considered as a weak DT. In this paper, we discuss the problem of DT classifica- tion based on optical flow. A recent brief survey of DT de- scription and recognition [3