Comparative Evaluation of Two Multisensory Video Surveillance Techniques for Pedestrian Tracking

In this paper we examine two different automated video surveillance techniques for detection and tracking of pedestrians based on fusion of colour and thermal images. The first approach is a novel particle filtering based on Bayesian framework, and the seconf one is an approach based on fusion of shape and appearance cues. The shape and appearance based technique involves a layered two pass scheme, where in the first pass-an expectation-maximization (EM) algorithm is used to separate infrared images into still background and moving foreground layers. In the second pass: shape cues from the first pass is used to eliminate non-pedestrian moving objects and then appearance cue is used to locate the exact position of pedestrians. Then pedestrians are detected by sequential application of templates at multiple scales. For tracking the pedestrian a graph matching-based algorithm which fueses the shape and appearance information was used. The particle filtering based algorithm on other hand is based on building a scene background model with each pixel represented as a multimodal distribution of colour and thermal images. Then this background model is used to build a particle filter for tracking the pedestrian. The particle filter uses a novel formulation of observation likelihoods The evaluation of the two detection and tracking approaches was done by performing experiments on the thermal and colour dataset from OTCBVS database