Research On Background-aware Correlation Filter Tracking

Visual object tracking is a hotspot in computer vision research due to its extensive applications and fundamental research significance.On the other hand,object tracking is a challenging vision perception task due to the universality of targets and the complexity of video scenes.Correlation filter based tracker(CFT)demonstrated the power in object tracking,benefiting from its superior speed and competitive performance.This paper focuses on unsolved issues in correlation filter tracking.CFT is built on the periodicity assumption of samples,and suffers from an intrinsic defect called the boundary effect.Both training and testing samples are generated by circularly shifting the base sample,which results in distortion and unreality of a high proportion of samples.The distortion constrains the model perceptivity and prediction capacity.In phase of model training,distorted samples affect the model's fitness on real video content;in the phase of model inference,the detection scores of boundary samples are unreliable,and cannot reflect the target confidence distribution on the boundary region.The ideas and contributions of this paper are summarized in three-fold.First,an in-depth analysis on the boundary effect reveals two significant unsolved problems of CFTs.One is the restricted model perception field and search range.This problem significantly affects tracking performance in various scenarios such as fast motion.Besides,after some temporary tracking failures caused by uncontrollable factors,a restricted search window makes it difficult to retrieve the target again,and results in unrecoverable tracking failures.The second problem is the data misalignment between training samples and testing samples.It is difficult to learn a high-quality model if testing samples and training samples severely violate the independent and identical data distribution.Second,a multi-anchor detection strategy and a distractor-aware filter are proposed to expand the target search field and model perception field,respectively.By constructing a pixellevel feature based Bayesian model or exploiting a Siamese network,we can determine some target-like regions efficiently in a wide context.Centered at target-like regions,multiple groups of circular samples are extracted for detection or training a distractor-aware filter.Multi-anchor detection achieves a wider search field and improves tracking performance in various scenarios such as fast motion,camera motion and occlusion,etc.The proposed distractor-aware filter learns from more background samples and enhances its background perceptivity.By specifically learning to suppress the response of distractors,the distractor-aware filter can reduce the risk of drifting to distractors during multi-anchor detection.Thirdly,based on the extended model perception field and search range,this paper proposes a progressive target search strategy of Proposal-Detect-Refine,and construct a novel tracking framework,namely DAMA.DAMA explores candidates in wider context via target proposal,and then extracts multiple groups of circular samples for distractor-aware filtering.Finally,a standard correlation filter is used for position refining.The progressive target search process gradually aligns the data distribution between testing samples and training samples,which benefits for reliable model prediction.Besides,the position refiner helps reduce error accumulation and prevents model from drifting.Comprehensive experiments on four datasets,i.e.,OTB2013,OTB2015,TC128 and VOT2017,demonstrated that DAMA achieves competitive performance with the state-of-theart deep methods,while maintains a real-time tracking speed.Besides,experimental results showed that DAMA achieves superior robustness in many challenging scenarios such as fast motion,camera motion,severe occlusion and out of view.