engleski

Structural analysis of video by histogram-based description of local space-time appearance

The methods for representation, classification and reasoning about video data are actively researched in the computer vision community, given a large number of interesting applications such as human action recognition, video segmentation, event detection, gesture recognition, and dynamic texture recognition. Different research applications drive different views on what is essentially the same type of information. The goal of this thesis is to investigate the methods for video analysis that are generally applicable to any kind of video data, regardless of the phenomenon represented by the data. Special emphasis is placed on the analysis of videos in an online scenario, where not all frames of a video are available in advance. The thesis introduces the notion of a spatio-temporal structure as a basic unit of video information. A method for representing spatio-temporal structures as grids of histogram is proposed, and the derived grid-of- histograms representation is used to build two different kinds of descriptors: the spatio- temporal appearance (STA) descriptors and the COIN descriptors. The STA descriptors model either the average local appearance or the distributions of local appearance, while the COIN descriptor models change in local appearance by assigning weights to four semantically meaningful hypotheses. A detailed experimental evaluation is provided, illustrating the suitability of STA descriptors as feature vectors in the problems of traffic sign recognition, human action recognition and dynamic texture recognition, and the suitability of COIN descriptors in determining the structural properties of the 3D world.

video analysis; spatio-temporal appearance; spatio-temporal structure; video descriptors; semantic descriptors; action recognition; dynamic textures

nije evidentirano