Evaluation of similarity measures for appearance-based multi-camera matching

Abstract

Visually matching people appearing in different camera views is an essential part of multi-camera tracking, camera hand-over and video-based identity search. The problem is made difficult by large variations in the appearance of subjects both within the same camera view and between cameras, as well as across time. Rather than relying on a single appearance-based matching method, a fusion of visual cues is more compelling. In this work 8 different similarity measures were evaluated encompassing shape, colour, texture and biometric information. The evaluation was performed on hand-labelled data from 4 indoor surveillance cameras. Experiments examined the accuracy of the similarity measures. Results revealed that matching accuracy is good when tracks come from the same camera, but poor when they come from different cameras. Although different measures performed best in different situations, the colour-based measure produced the best results overall.