Adaptive Tracking via Multiple Appearance Models and Multiple Linear Searches

Abstract

This research is concerned with adaptive, probabilistic single target tracking algorithms. Though visual tracking methods have seen significant improvement, sustained ability to capture appearance changes and precisely locate the target during complex and unexpected motion remains an open problem. Three novel tracking mechanisms are proposed to address these challenges. The first is a Particle Filter based Markov Chain Monte Carlo method with sampled appearances (MCMC-SA). This adapts to changes in target appearance by combining two popular generative models: templates and histograms, maintaining multiple instances of each in an appearance pool. The proposed tracker automatically switches between models in response to variations in target appearance, exploiting the strengths of each model component. New models are added, automatically, as necessary. The second is a Particle Filter based Markov Chain Monte Carlo method with motion direction sampling, from which are derived two variations: motion sampling using a fixed direction of the centroid of all features detected (FMCMC-C) and motion sampling using kernel density estimation of direction (FMCMC-S). This utilises sparse estimates of motion direction derived from local features detected from the target. The tracker captures complex target motions efficiently using only simple components. The third tracking algorithm considered here combines these above methods to improve target localisation. This tracker comprises multiple motion and appearance models (FMCMC-MM) and automatically selects an appropriate motion and appearance model for tracking. The effectiveness of all three tracking algorithms is demonstrated using a variety of challenging video sequences. Results show that these methods considerably improve tracking performance when compared with state of the art appearance-based tracking frameworks.