Research Of Visual Object Tracking Based On Correlation Filter

Visual object tracking is one of the most challenging tasks in computer vision,and it plays a crucial role in many applications such as video surveillance,video retrieval and precision guidance.In order to explore powerful feature representation for complicated tracking scenarios,numerous tracking algorithms have been proposed,which range from traditional methods based on object matching to current methods based on learning.In these methods,correlation filter-based tracking algorithm is attracting great attention because of the good performance and high efficiency.In this paper,we propose several improved algorithms based on the analysis and understanding of correlation filter.The main research contents are as follows:(1)To address the shortcomings of existing correlation filter-based trackers in feature representation,an adaptive multi-feature fusion method is proposed on the basis of kernel correlation filter(KCF)framework.Firstly,multiple features are introduced for fusion to obtain more complementary and powerful feature description.Secondly,a hybrid multi-feature fusion strategy is proposed and the fusion weight is calculated adaptively.(2)To effectively select informative features,an adaptive feature selection method is presented on the basis of KCF framework.Firstly,three complementary features are extracted to learn three independent filter models,and then by analyzing the relationship between response maps and tracking results,the most representative feature is selected adaptively for tracking.Secondly,the model update strategy is further improved to better handle occlusions and drifts.(3)On the basis of above improvements on feature representation,a separate scale correlation filter is further introduced for scale estimation.We present two tracking algorithms,namely multi-scale correlation filter tracker based on adaptive feature fusion and multi-scale correlation filter tracker based on adaptive feature selection,which improve the scale adaptability and generalization ability of tracking algorithms.We evaluate our proposed algorithms using visual tracker benchmark with 51 videos.The experimental results show that compared with the traditional correlation filter tracker and other state-of-the-art trackers,our algorithms not only achieve higher accuracy,but also guarantee the real-time tracking speed.