Robust Likelihood Model for Illumination Invariance in Particle Filtering

Abstract

Tracking visual targets in an unconstrained environment is challenging due to variations in illumination, scale, occlusion, and motion blur, for example. Many video applications that utilize particle filter-based visual target trackers require tracking of visual targets under varying illuminations. Similarity measures and likelihood estimation strongly influence the performance of particle filters. In this paper, we propose a novel likelihood estimator that has been combined with other state-of-the-art particle filtering-based tracking techniques to accommodate varying illumination by predicting changes in the illumination intensity and direction of the illumination. Moreover, an enhanced update strategy for the template dictionary is used along with a sparse representation model to solve the problem of drift due to appearance changes during tracking. The proposed algorithm has been evaluated using various particle-filter-based tracking algorithms on scenes from public data sets and using our gesture data set, which includes variations in illumination. Using the proposed model, the algorithms perform up to 20% better on sequences for which variations in illumination are dominant. We carried systematic experiments to evaluate the robustness of the proposed algorithm on video sequences with illumination variations, as well as other variations. Furthermore, in sequences that include variations in illumination, our likelihood model usually performs better than the default tracker likelihood model.