Object Tracking and Mensuration in Surveillance Videos

Abstract

This thesis focuses on tracking and mensuration in surveillance videos. The

first part of the thesis discusses several object tracking approaches based on the

different properties of tracking targets. For airborne videos, where the targets are

usually small and with low resolutions, an approach of building motion models for

foreground/background proposed in which the foreground target is simplified as a

rigid object. For relatively high resolution targets, the non-rigid models are applied.

An active contour-based algorithm has been introduced. The algorithm is based on

decomposing the tracking into three parts: estimate the affine transform parameters

between successive frames using particle filters; detect the contour deformation using

a probabilistic deformation map, and regulate the deformation by projecting the

updated model onto a trained shape subspace. The active appearance Markov chain

(AAMC). It integrates a statistical model of shape, appearance and motion. In the

AAMC model, a Markov chain represents the switching of motion phases (poses),

and several pairwise active appearance model (P-AAM) components characterize the

shape, appearance and motion information for different motion phases. The second

part of the thesis covers video mensuration, in which we have proposed a heightmeasuring

algorithm with less human supervision, more flexibility and improved

robustness. From videos acquired by an uncalibrated stationary camera, we first

recover the vanishing line and the vertical point of the scene. We then apply a single

view mensuration algorithm to each of the frames to obtain height measurements.

Finally, using the LMedS as the cost function and the Robbins-Monro stochastic

approximation (RMSA) technique to obtain the optimal estimate.