Improving target tracking robustness with Bayesian data fusion

Abstract

Intelligent data fusion is an active area of research. Most recent works in data fusion for object tracking employ machine learning techniques that lack flexibility due to their inability to adapt to changing conditions in the presence of limited amounts of training data. Our work explores a hierarchical Bayesian fusion approach, which aggregates information from multiple tracking algorithms into a more robust estimate and hence outperforms its constituent trackers. This adaptive and general data fusion scheme takes advantage of each tracker's local statistics and combines them using a global softened majority voting. The widespread availability of high-performance multicore processors has allowed parallel threads to run multiple trackers asynchronously, which means that the algorithm can be executed in real time as it is only limited by the slowest tracker in the ensemble. The proposed approach is corroborated and evaluated on the OTB-50 dataset.