Spherical Vision for Visual Surveillance

点击翻译按钮获取中文摘要

A recent analysis of the needs and challenges of intelligent video surveillance has shown that COTS systems lack many useful video processing features because academic research has not produced reliable results for most real-world applications. Therefore, state-of-the-art commercial systems only rely on basic image processing and do not adequately support security agents. Video processing features urgently needed for enhance security and situation awareness include functions for human tracking (to localize a person at each instant for identification) and understanding the activities in the scene (to detect unusual behaviors or specific actions like running). These two functions are usually investigated using standard fix cameras, although PTZ cameras were used in some recent research. Single fix cameras cover small areas of visibility in surveillance application. The use of cameras that can cover a larger area is often considered (PTZ, and more recently spherical). Our previous research with PTZ camera has shown the difficulty with handling changing viewpoints during person tracking. Indeed, people are usually viewed from the same angle with a fix camera. This is not true with PTZ cameras as they rotate (the view can change from side to frontal), nor the new spherical cameras. Because of the nature of a spherical camera, a person crossing a scene can be at first viewed from the front, at some point from the top, and then from the side or the back. With a spherical vision camera, we aim to study in more detail the issue of viewpoint change in the context of person tracking and recognition in video surveillance. Our initial application for the camera will be the recognition of individuals for person identification from a watchlist under very large viewpoint changes. We will first track the people in the scene to collect information about their appearance, and then progressively match the appearance of each person against the watchlist using ensemble of classifiers. More globally, our objectives are (1) to develop robust tracking and recognition techniques to handle the extreme viewpoint change provided by spherical cameras, and (2) to investigate strategies to model human activities under these changing viewpoints.