Object Tracking based on KCF and Sparse Prototypes

Recently, many correlation filter-based tracking methods have received lots of attention and achieved great success in visual object tracking. Among correlation filter-based methods, the most influential one is kernelized correlation filter (KCF) which has excellent performance both in efficiency and accuracy. However, due to the virtual nature of cyclic shifts samples, the training and detecting of KCF are imprecise. To alleviate the influence of virtual samples, we take the following two measures. 1) We extract image patches at positions of samples which have local maximum KCF responses and treat them as candidates. We further evaluate their true responses. 2) Instead of estimate the target position directly according to the KCF results, we use the sparse prototypes (SP) as the target model to evaluate the similarities between candidates and target. The results of KCF and SP are combined by adaptive weight to estimate the target position. In addition, KCF degrades due to its unreasonable update scheme. To do reliable update, we set different update modes and generate an adaptive update rate based on two tracking confidence indices. Experiments on a commonly used tracking benchmark show that the proposed method improves KCF about 8% on the average success rate and 10% on the precision, and achieves better performance than other state-of-the-art trackers.