Real time target tracking with a single visual sensor

This dissertation presents a scientific approach to real time target tracking using a single visual sensor. The objective is to formulate strategies to track N targets with stochastic motions by sharing a single visual sensor.;Allocating the visual sensor to a target just enough to maintain successful tracking will ease the restriction of continuously watching the desired target; the key to successful tracking is to identify the proper sampling rate for each target.;One characteristic of visual tracking systems is the rather long processing time associated with extracting information from the image. This dissertation presents methods to reduce the processing time by utilizing the target motion dynamics; instead of processing the whole image, a Region of Interest (ROI) is processed to locate the target of interest.;Due to the stochastic nature of the targets' dynamics, the search window (ROI) within the sampled image is non-deterministic and its size can vary unpredictably; hence, the feed back rate is erratic and the tracking outcome becomes uncertain. A successful, stable tracking is only achievable if the target dynamic states are updated sufficiently through a reliable, deterministic sampling rate.;In multitarget tracking case, algorithms are developed to guarantee that each target will receive the minimum feed-back rate needed to sustain stable tracking. Target scheduling schemes are constructed to organize the order in which each target is served while meeting its temporal constraint. The limited processing time allocated to each target is then used to process the search window; instead of processing all pixels within the search window, only a handful of pixels are processed.;In this thesis, the Adaptive Resolution Control (ARC) mechanism is introduced to limit the processing time spent on each sampled image and ensure delivery of a minimum feed-back rate so vital for achieving successful tracking. The ARC deploys a multi-stage processing mechanism in which the sampled image is processed in sub-resolution in first stage and higher resolution when it finds a trace of the intended target.