Motion Trajectory-Based Video Retrieval, Classification, and Summarization

Abstract

This chapter provides an overview of various methods for motion trajectory-based video contentmodeling, retrieval and classification. The techniques discussed form the foundation for content-based video indexing and retrieval (CBVIR) systems. We focus on view-invariant representations of single and multiple motion trajectories based on null-space invariants that allows for video retrieval and classification from unknown and moving camera views. We introduce methods based on matrix and tensor decomposition for efficient storage and retrieval of single and multiple motion trajectories, respectively. We subsequently explore the use of one- and multi-dimensional hidden Markov models for video classification and recognition based on single and multiple motion trajectories. We summarize the basic concepts and present computer simulation results to demonstrate the fundamental notions introduced throughout the chapter.We finally discuss several open problems in the field of motion trajectory analysis and future trends in content-based video modeling, retrieval and classification.