Pedestrian detection and re-identification in surveillance video

The detection and re-identification of pedestrians is an important component of the automated analysis of surveillance video. The main challenge addressed herein is the development of accurate and efficient algorithms for these two tasks, using oversampled. redundant sets of features to which machine learning algorithms may be applied. For pedestrian detection, Integral Line Scan Features (ILF) are developed as a means of generating such a pool of features. A machine learning section procedure can be used to derive an appropriately weighted subset of features. One advantage provided by the integral design is that dense sampling of the image is relatively efficient, since the integral features are calculated only once for the entire image. Another advantage provided by the feature characteristics is a relatively consistent performance across different feature scales, which obviates the need to generate a scaled image pyramid. Methods for pedestrian re-identification, using redundant feature sets, are also investigated. It is hypothesised that performance can be improved by simultaneously expressing features in two alternative colour spaces, using both to learn a transformation into a metric space well suited to the task. Experiments are presented that confirm this hypothesis, using a novel adaptation of the Local Fisher (LF) machine learning approach. A separate contribution to the re-identification problem is the development of a method, SELF : that uses multiple classifiers. Each classifier is assigned to a given category of difficulty (of re-identification). It was hypothesised that such a method would improve performance, and experiments were devised to verify this idea. A final contribution is the analysis of pedestrian re-identification performance metric from an information-theoretic perspective, and the proposal for a metric that measures the proportion of uncertainty (PUR) removed. This metric can be applied to represent pedestrian re-identification performance. The thesis concludes with a discussion of implications and future extensions