Research On Object Tracking Algorithm Based On Improved Correlation Filter With Multifeature Fusion

Object tracking is always a challenging subject in computer vision.However,most online learning tracking algorithms will encounter some problems since the target may undergo major morphological changes in the process of tracking and lack sufficient training samples.For example,when the target is occluded,out of view or fast-moving,it is easy to cause model drift and results in tracking failure.In this thesis,an improved algorithm based on correlation filters is proposed for the object tracking task in complex scenarios.The performance is boosted by a combination of several strategies termed multi-feature fusion,multi-scale estimation,adaptive model update and re-detection after target loss.The main research objects,content and innovations are as follows:(1)Aiming at the problem of various target forms and dramatic changes in appearance during the tracking process,an improved adaptive multi-feature fusion tracking algorithm is designed.Firstly,the depth features are extracted from the target search area by a convolutional neural network,and the specific multi-layer convolutional results are weighted and fused according to the target information contained in the convolutional layers of different depths in the neural network.Secondly,two manual features,HOG and CN,are extracted from the search area and merged in a cascading way.Finally,the response maps calculated by depth feature and manual feature are adaptively fused according to their confidence,and the fusion response map is used to calculate the target position.Consequently,the advantages of multiple features can be combined to overcome the problem of tracking instability caused by the dramatic changes in the appearance of the target in the process of tracking.(2)In order to solve the problem that the target scale changes by rotation and deformation,an improved multi-scale target estimation method is designed.The method of extracting target candidate bounding box proposals through Edge Box is combined with the traditional scale pyramid estimation method,and the optimal result of the two is chosen as the final target estimation scale,thus avoiding the problem that the aspect ratio of target bounding box is always unchanged when using the single scale pyramid method for scale estimation,which significantly improves the expressive ability of target.(3)For the problem of model drift even lost by the occlusion or out of view of the target,an improved adaptive model update as well as target loss re-detection tracking algorithm is proposed.Firstly,tracking results are evaluated by the peak value and the average peak-to correlation energy indicators of the response map.If the target is judged as occlusion,the updating process is suspended to avoid drifting.Secondly,a logistic regression classifier and a long-term correlation filters model are trained in the tracking process.When the target is lost,the Edge Box algorithm is adpoted to extract target candidate sample proposals in the surrounding area,then the trained classifier and model is utilized to simulate the candidate proposals,the optimal proposal is regarded as the re-detection target position,so as to recover the lost target and improve the robustness of tracking algorithm.Finally,the proposed algorithm is tested and analyzed on OTB and VOT datasets.Experimental results with other state-of-the-art tracking algorithms prove the effectiveness and advantages of the proposed algorithm.