Research On Object Tracking Based On Correlation Filtering Framework

Target tracking technology is closely related to our daily life and production.It has great application value in many fields such as civilian and military.However,there are many factors affecting tracking performance.Improving the robustness of tracking system is still facing severe challenges.At present,correlation filtering is one of the most important methods in the field of target tracking,but in some complex tracking environments,this kind of algorithm still needs to be further studied.Therefore,based on the correlation filtering framework,this thesis studies the improved methods of target tracking from various perspectives.The main work and achievements include the following:(1)A correlation filtering tracking algorithm combined with particle filter is studied.Firstly,correlation filtering is used to track the video sequence.The peak sidelobe ratio of the filter response graph is used to determine whether occlusion occurs in the current frame.If no occlusion occurs,the target position predicted by correlation filtering algorithm is the final tracking result.If occlusion occurs,a tracking scheme combining correlation filtering and particle filtering is adopted.For the scheme of combining correlation filter and particle filter,firstly,the transfer model is used to generate particles and initialize them;secondly,the particles are randomly sampled from the importance probability density function;thirdly,the weights of particles are updated and resampled;then,the resampled particles are filtered by correlation filter,and the weights of each particle are modified according to the response of the correlation filter to realize the selection of particles;finally,the weighted average is calculated to get the particles' state.Relevant filtering can guide particles to move towards the target distribution mode,reduce the number of particles to a certain extent,and reduce the computational complexity of particle filtering.(2)A correlation filtering tracking algorithm combining multi-channel features is studied.Three independent correlation filters are trained by the complementary features of image shape,color and texture.According to the response of correlation filter,different weights are set for each independent filter.The weighted response graph is obtained by linear superposition.The position of the largest response value in the weighted response graph is the final target position.In order to solve the problem of scale change,the algorithm also adds a scale estimation module.By setting scale pools,feature extraction is carried out for samples of different sizes,learning a scale filter,and estimating the target scale with a scale filter.Because the scale estimation module is added to the algorithm,it can zoom the target of different sizes to a certain extent,which solves the problem of tracking the target scale fixed in the correlation filtering algorithm.(3)A correlation filtering tracking algorithm combining LK optical flow(Lucas-Kanade,LK)is studied.Firstly,forward and backward optical flow matching is performed on the tracking video sequence,and the number of feature points after matching is used to determine whether the target is occluded or not.If occlusion occurs,the target prediction position is obtained by optical flow tracking,because optical flow carries abundant motion information and structure information of moving objects,and can detect moving objects in unknown scenes.If occlusion does not occur,the target prediction position is obtained by correlation filtering algorithm with high operational efficiency.In addition,after determining the prediction position of the target,the scale change ratio of the target can be calculated by using the selected key feature points,and the adaptive change of the target scale in the current frame can be realized.The above algorithm is validated on OTB2013 data set.The experimental results show that the correlation filtering tracking algorithm combined with particle filter can effectively solve the partial occlusion problem.Because particle filter can adapt to the change of target scale,the proposed algorithm can also solve the problem of target scale change to a certain extent,but there is still room for improvement in tracking speed.The correlation filtering tracking algorithm combined with multi-channel features can represent the target more accurately,which improves the tracking accuracy in complex scenes such as background clutter and illumination change.Scale estimation module is added to the algorithm,which can solve the problem of target scale change in a certain range and maintain a relatively high tracking speed.The correlation filter tracking algorithm combined with LK optical flow can effectively deal with various occlusion problems.At the same time,for the target scale change problem,the scale change ratio is not limited to a certain range,but can realize the scale adaptive adjustment.