Alexander, Daniel; (1997) Statistical Modelling of Colour Data and Model Selection for Region Tracking. Doctoral thesis (Ph.D), UCL (University College London).

Text Statistical_modelling_of_colou.pdf Download (26MB)

Abstract

This thesis is concerned with the question of how best to model naturally occurring distributions of digital colour camera data so that objects may be characterised in terms of such data and thus tracked through sequences of images. By use of a physical model of the processes by which colour data is obtained, several statistical models, based on directional statistics, are proposed as alternatives to the standard multivariate Gaussian distribution. Various types of algorithm are discussed that can use these statistical colour models for tracking regions of interest through sequences of images. One such region tracking technique, the active region model or statistical snake, is investigated in detail and some improvements to the original model are suggested. In order to decide which of a set of candidate statistical models and which of a set of region tracking algorithms is the best, a methodology for measuring the performance of these algorithms in terms of pixel classification is developed. This methodology is used to optimise the internal parameter settings of several region tracking algorithms and to compare their overall performance. For the connected, compact regions of interest in the test set, active region models are shown to outperform simpler thresholding and region growing algorithms in terms of pixel misclassifications. Some of the suggested alterations to the existing active region model implementations are shown to increase performance. However, it is also found that the introduction of some desirable properties, such as stability and convergence, can increase susceptibility to local minima and reduce overall classification performance. The set of candidate SCMs are compared over two data sets, representative of two separate applications, by comparing the performance of region tracking algorithms using each candidate model. It is shown that in controlled environments directional models are superior to Gaussian models. However, for imagery obtained in less constrained daylight environments, generalisation of the directional models is found to be poor and the Gaussian model proves to be preferable.

Download activity - last month

Export as JSONXMLCSV

Download activity - last 12 months

Export as JSONCSVXML

Downloads by country - last 12 months

Export as JSONCSVXML

Archive Staff Only

View Item