Performance Evaluation of Object Representations in Mean Shift Tracking
Peter Hosten, Andreas Steiger, Christian Feldmann, Christopher Bulla · RWTH Publications (RWTH Aachen) · 2013
Mean shift tracking is a real-time capable object tracking approach that is not restricted to a specific object category. Several target object representations based on a feature distribution within an object region have been proposed for mean shift tracking. Quantitative performance metrics for the evaluation of object representations in mean shift tracking are mainly based on a comparison against ground truth data, which is often not available or requires considerable effort for its creation. In this paper, our main contribution is a novel approach for the quantitative evaluation of object representations in mean shift tracking, that does not rely on any ground truth data. Our approach is based on multiple hypotheses for the object location which initialise the mean shift tracking algorithm. The tracking result is then treated as random process and a quantitative metric is derived from its properties. Finally, the evaluation approach is applied to various object representations and test sequences. The findings demonstrate that the usage of multi-part object representations is beneficial if the representation captures the spatial colour distribution of the object. Keywordsmean shift tracking; multi-part object representation; tracking evaluation