A Comparative Study of Boosted and Adaptive Particle Filters for Affine-Invariant Target Detection and Tracking

This paper addresses a specific multi-aspect target detection and tracking problem where the dynamics of the target’s aspect is modeled by an affine model following a first-order Markov process. We are interested in how to achieve robust and accurate Monte Carlo estimation in a high-dimensional state space with poor target visibility by re-visiting two recent improvements to particle filters, i.e., "boosting" and "adapting". The impetus of this work is a tracking indicator that estimates the tracking performance based on the observation model and may trigger either one of two actions when it is necessary. One is "boosting", i.e., the detector specified by the tracker’s previous output is involved to induce more promising particles, and the original idea of "boosting" is extended here by encouraging positive interaction between the detector and the tracker. The other is "adapting", i.e., the system model can self-adjust to enhance the tracking capability. We compare two methods in the context of affine-invariant target tracking and with respect to their contributions to improve the particle quality. Experiments on simulated image sequences with real infrared background show that both techniques can improve the tracking performance by balancing the focus and the diversity of particle distribution.