Research On Regularized Adaptive Object Tracking Based On Particle Filtering

Visual object tracking,as the main branch of computer vision field,has a good and mature development community.While there have been good improvements in tracking accuracy and robustness in recent years due to the proposed correlation filtering methods,research on efficient tracking algorithms that cope with as many challenging factors as possible continues to receive sustained attention due to the complex changes in real-life scenarios.The main reasons affecting the performance of the tracking algorithm are: first,the uncertainty brought by the rectangular box initialization to the tracking algorithm and the limitation of the searching area,where the foreground and background information can’t be used more effectively;second,the inevitable introduction of the boundary effect problem when training the filtering template by the kernelized correlation filtering method,where the real target representation is affected;third,the model degradation problem accumulated during continuous tracking due to occlusion and drastic deformation.To address the above problems,in this research,a correlation filtering method to train the model under the framework of particle filtering is introduced,and the main research contents are as follows:(1)Research on kernelized correlation filtering tracking algorithm based on particle filtering.To address the initialization problem at the beginning of tracking and the limitation of the searching area during tracking,the particle filtering method is combined with the correlation filtering method,and in the initial frame,multiple particle blocks of different sizes are generated near the center position of the target,which is not limited to the initialization of a single random bounding box,and the model is trained using the kernelized correlation filtering method after traditional features are extracted in a more efficient searching area,and the final position and scale size of the target are obtained based on the particle trackability and particle motion similarity to determine reliable particle blocks.(2)Research on tracking algorithms via spatio-temporal regularization.To address the boundary effect problem arising when training the target tracking model,the temporal regularization term and the spatial regularization term are introduced,the training data is increased by using the cyclic shift structure of each frame sample,the filter coefficients are penalized in the training session,which makes the filter pay more attention to the real target information,and the model obtained from the historical frames is considered in the training paradigm to limit the drastic changes of the filtering model,where a large degree of improvement in both accuracy and robustness are obtained.(3)Research on model adaptive tracking algorithm based on particle filtering.To address the problem of model degradation caused by occlusion etc.in the tracking process,the causes of tracking loss due to model degradation is analyzed.A mean peak-to-fluctuation ratio of confidence map indicator is proposed,quantifying the response score plot,and the temporal regularization term is adaptively adjusted,then compare it with the historical frame average of the indicator to measure the reliability of tracking results,so as to further determine the probability of updating the model.The model adaptive tracking algorithm proposed in this study is further applied to multiple particles under the particle filtering framework,where better results can be obtained in dealing with the model drift and degradation problems caused by challenges such as occlusion and the accuracy of target state assessment is improved.