Representing and Modeling Human Actions in Videos.

摘要

Human action analysis in videos is an important and fundamental problem in computer vision due to its potential applications, such as intelligent analysis of surveillance videos, patient monitoring systems, humancomputer interfaces, content based video retrieval, shot detection in movies and sports videos. In this thesis, we study the problem of human action analysis, in particular with a focus on three aspects of this problem, namely (i) seeking mid-level "parts" for video representation, (ii) designing compositional models for detailed action understanding, and (iii) applying deep networks to action recognition. (Abstract shortened by ProQuest.).