A Superior Tracking Approach: Building a strong Tracker through Fusion
Christian Bailer, Alain Pagani, Didier Stricker
Computer Vision – ECCV 2014 European Conference on Computer Vision (ECCV-14), September 6-12, Zürich, Switzerland

Abstract:

General object tracking is a challenging problem, where each tracking algorithm performs well on different sequences. This is because each of them has different strengths and weaknesses. We show that this fact can be utilized to create a fusion approach that clearly outperforms the best tracking algorithms in tracking performance. Thanks to dynamic programming based trajectory optimization we cannot only outperform tracking algorithms in accuracy but also in other important aspects like trajectory continuity and smoothness. Our fusion approach is very generic as it only requires frame-based tracking results in form of the object’s bounding box as input and thus can work with arbitrary tracking algorithms. It is also suited for live tracking. We evaluated our approach using 29 different algorithms on 51 sequences and show the superiority of our approach compared to state-of-the-art tracking methods.

Keywords:

Object Tracking, Data Fusion

A Superior Tracking Approach: Building a strong Tracker through Fusion
(Hrsg.)
Computer Vision – ECCV 2014 European Conference on Computer Vision (ECCV-14), September 6-12, Zürich, Switzerland

Abstract:

General object tracking is a challenging problem, where each tracking algorithm performs well on different sequences. This is because each of them has different strengths and weaknesses. We show that this fact can be utilized to create a fusion approach that clearly outperforms the best tracking algorithms in tracking performance. Thanks to dynamic programming based trajectory optimization we cannot only outperform tracking algorithms in accuracy but also in other important aspects like trajectory continuity and smoothness. Our fusion approach is very generic as it only requires frame-based tracking results in form of the object’s bounding box as input and thus can work with arbitrary tracking algorithms. It is also suited for live tracking. We evaluated our approach using 29 different algorithms on 51 sequences and show the superiority of our approach compared to state-of-the-art tracking methods.

Keywords:

Object Tracking, Data Fusion