Soft-assigned bag of features tracking

Bag of features (BoF) provides an effective and efficient representation for object tracking in video sequences. However, hard assignment used in BoF generation inevitably brings in quantization errors, which may lead to inaccuracy even failure in tracking. In this paper, we propose a novel soft-assigned bag of features tracking approach (SABoF), which can significantly reduce the influence of quantization errors and obtain more accurate and stable tracking results. We initialize tracking by specifying the tracked object and constructing the codebook. Then, we represent each candidate target with soft-assigned BoF and measure its similarity to the tracked object. The most similar candidate target in each frame is selected as the tracked result. To improve tracking performance, we also refine the tracking results by combining incremental PCA tracking. The proposed approach is evaluated on the challenging video sequences from CAVIAR dataset. Experiments show our approach outperforms current dominant methods in complex conditions.