Research On Feature Selection And Correlation Filter-based RGBT Tracking

Visible light(RGB)imaging reflects the color and texture characteristics of the object,and its appearance can be recognized well under good illumination.While the thermal infrared(T)imaging reflects the temperature distribution of the object surface and is insensitive to light and can penetrate the smoke well,therefore,while can better overcome the influence of haze and other bad environments.In recent years,visible-thermal infrared(RGBT)tracking has attracted wide attention due to the complementary advantages of RGB and T imaging.At present,most researchers pay attention on the complementarity between visible light and thermal infrared image,and strengthen the robustness of target tracking through modal fusion.However,many RGBT tracking methods integrate RGB and T modalities into a whole for tracking and do not consider that the reliability of the modality is likely to change alternately with time,and the modality with poor reliability will reduce the tracking performance.In addition,the complex tracking model has high computational cost,which makes the tracking speed be non-real time.To solve the above problems,this thesis studies the selection of reliable features under the framework of efficient correlation filter,which mainly includes two aspects as follows:First,this thesis proposes an RGBT tracking algorithm based on feature selection in the framework of correlation filter since the reliability of RGB modality and T modality changes with time in the process of target tracking.Through dynamic feature selection,the appearance model in the tracking process is more reliable.Firstly,this thesis uses RGB image,T image and RGBT image to extract and combine a variety of features.Secondly,a criterion combining consistency,smoothness and robustness is proposed to evaluate the trackers with different feature,so it can dynamically select the most reliable tracker.Finally,the experimental results on two large-scale RGBT tracking datasets show that the proposed method makes good use of the complementarity of the two modalities and achieves a more accurate tracking result.Second,aiming at the difference of feature channels and tracking efficiency,this thesis dynamically learns the reliability of each feature channel,and uses principal component analysis to reduce the dimension of features in order to improve the tracking speed.Under the framework of correlation filter,a real-time RGBT tracking algorithm based on weighted feature channels is proposed.Firstly,the feature dimension of RGBT position estimation is reduced by trigonometric interpolation and principal component analysis.In the process of scale estimation,fewer scale estimations are selected to obtain more scale responses,which can greatly improve the tracking efficiency.Secondly,the filter is weighted by learning and detecting channel reliability,and the most robust tracker is selected to complete the tracking.Finally,experiments on two large-scale RGBT tracking datasets prove that the method achieves a real-time tracking efficiency and good tracking performance.