Research On Stable Correlation Filter Algorithm For Visual Tracking

As one of the important issues in the field of computer vision,target tracking has been widely applied in video surveillance,unmanned driving and human-machine interaction.Due to the complexity and variety of tracking scenarios,the target may be affected by changes in its own motion patterns or external environmental factors.It is still very difficult to develop a robust tracking algorithm that can deal with various scenes.This thesis combines the correlation filter framework and the characterization ability of convolution features.The existing algorithm is improved from the aspect of exploring reliable model and robust features.Experimental results show that the accuracy and stability are enhanced while maintaining advantage in speed.The main research contents are as follows:Based on correlation filter,an algorithm named ACHU was proposed to improve the Staple algorithm in terms of model which utilizes adaptive context awareness.The key point is to optimize the structure of samples and adjust the update strategy.To solve the problem of boundary effect caused by cyclic matrix,the number of reasonable samples is increased by integrating the real background information of the context as training samples.Then Kalman filter is employed to estimate the motion direction of the target.During training procedure,the context samples in this direction are adaptively assigned a larger weight,which enhances discrimination ability of the model and retains a closed solution.In addition,aiming the problem of adopting fixed learning rate,the indicator named APCE is introduced to judge the confidence of the response graph.Based on this,a linear update strategy is proposed,which can avoid the model pollution at the time of target occlusion.From the aspect of features,the ACHU algorithm is further improved.Based on the strong representation and generalization ability of convolution features,an algorithm named Deep ACHU is proposed to exploit hierarchical convolution features.Visual experiments show that the high resolution of shallow convolution features is conducive to accurate positioning,while the semantic information contained in deep convolution features is helpful for classification.The offline-trained VGG-16 network is selected as a feature extraction model,and the feature maps output by the two convolution layers are used for training the correlation filters after bilinear interpolation.Due to the overlap between the context and the target area,receptive field mapping is used to speed up the process of feature extraction.Finally,the detection response graphs of different models are used to construct reliable weight coefficients based on the primary and secondary peak ratios,and the models are fused at the decision level.The proposed algorithm ACHU and Deep ACHU are tested on public datasets OTB-2013 and OTB-100,also compared with some leading approaches.Quantitative and qualitative results fully prove its effectiveness and robustness.