LAUSR

Correlation-filter-based trackers, showing strong discrimination ability in challenging situations, have recently achieved superior performance in visual tracking. However, because the model treats the tracker's predictions in new frames as training data, the filter can be contaminated by small incorrect predictions, which cause model drift. Particle-filter-based trackers usually produce more accurate results due to the richer image representations used in prediction, but suffer when the environments are complex throughout an image sequence. In this letter, we propose an innovative real-time algorithm, which combines the particle filter with correlation filters in the prediction stage, enabling accurate predictions by the particle filter and alleviating model drift. Moreover, an effective decision fusion strategy is proposed to get more precise object predictions, thus further enhancing the overall tracking performance. Extensive evaluations on the OTB-2013 benchmark demonstrate that the proposed tracker is very promising compared with the state-of-the-art trackers, while operating over 85 frames/s.