Target Tracking Based Scene Analysis

摘要

Target Tracking and Scene Analysis are two research areas in Computer Vision which, in the past, have been dealt with as separate problems. Most target tracking systems have the relatively modest goal of detecting and identifying moving objects and tracking them, using prediction and correlation techniques, as long as they move unobscured across the field of view. Most of these systems have been developed with a set of stringent real time constraints which, given current hardware technology, constrains them to employ a set of computationally simple algorithms. Most scene analysis/segmentation systems analyze a given scene based on the static properties of objects such as shape/structure and texture, where the class of objects are generally predefined. These systems tend to be slow whenever the scene is complex, containing many objects. There are advantages to combining the two problems into one. This paper is based on such and idea is concerned with the design of an Intelligent Target Tracking System (ITTS). This experimental system tracks targets based not only on models for targets (shape, motion, etc) but also on models of the environment through which the targets navigate and of the sensing system(s) employed to acquire the time-varying images on which the analysis is based. The ITTS is composed of a number of processes which can be divided into three processing stages: target recognition, segmentation of time-varying images, and scene model generation.