Static and dynamic human analysis in images and videos.

摘要

The purpose of the work contained in this dissertation is to propose and study a set of tools and techniques with which human activities in images and video can be extracted and analyzed. The significance of the work stems from the ever-growing needs of automation and machine understanding in areas ranging from video database indexing to practical surveillance systems. Several fundamental problems in computer vision communities (e.g., accurate image-based object detection/recognition, robust object tracking, and motion event detection) are addressed in this dissertation. The solutions of these problems can serve as the stepping stones to build other higher-level computer-human intelligent interfaces.; The dissertation encompasses studies on human face and body modeling, face detection and face recognition in images, tracking feature selection, human tracking in videos, human motion event classification in videos, and multimedia video retrievals. The key theme of this dissertation is the word search. We search for novel features for detection, recognition, tracking and motion events. We search for efficient ways to explore vast data space in multimodal and multimodel tracking and event detection scenarios. We also strive to find effective way to search for content-relevant material in large video databases.