Research On Feature Fusion Based Correlation Filter Tracking

Visual object tracking is one of the most challenging tasks in computer vision.It plays a crucial role in many applications such as human-computer interaction,video surveillance,and unmanned vehicle.Recently,the correlation filter-based trackers(CFTs)have aroused increasing interests because of its superior performance on both computational efficiency and tracking accuracy.In recent multi-feature fusion based CFTs,the efficiency of feature fusion is relatively low and causes poor real-time performance.In view of this shortcoming,this paper mainly focuses on improving the efficiency of feature representation and puts forward a series of innovative methods.The main research contents are as follows:(1)Starting with theoretical basis of correlation filters,two main criteria are used to evaluate the description capability and complementarity of different features in multiple tracking scenarios.The evaluation results further provide theoritical and experimental basis for feature selection in following researches.(2)In order to overcome the shortcomings and limitations of the traditional serial feature fusion strategy,the Principal Component Analysis(PCA)method is proposed to compress the feature redundancy in the serial fused feature.The PCA projection can increase the density of effective feature representation and improve the feature fusion efficiency so as to enhance the accuracy and robustness of the tracking algorithm.(3)According to the representation model of human vision system,this paper proposes a progressive tracking framework,which can adaptively choose the more efficient feature representation for both ordinary and complex tracking scenarios.Thus the tracking process can be further accelerated while still maintain good robustness.By using a variety of standard tracking data sets and evaluation methods,the proposed algorithms have been extensively validated.Experiments show that,compared with the existing correlation filter tracking algorithms and other classic tracking algorithms,the proposed algorithms achieve the state-of-the-art performance in both accuracy and robostness when performing the tracking at approx.50 frames per second.