Visual Object Tracking Based On Correlation Filter And Hierarchical Weight

Visual tracking is one of the core problems of computer vision.Its main task is to detect moving targets from successive video image sequences,and then determines the position of the moving target in each frame of images.Object tracking has wide-ranging applications,including driverless vehicle,intelligence video surveillance,human-computer interaction,and medical image tracking analysis.However,due to objective factors such as partial and full occlusion,background clutters,geometric deformation,illumination on changes,and viewpoint variation,to name a few,achieving accurate object tracking remains a challenging task.In this paper,we have conducted in-depth analysis and researched on objective factors such as fast motion,partial and full occlusion,and background clutters.The main contributions of this work as follows:1.We present an object tracking method which is based on observation likelihood model.In order to distinguish the effects of the patch-trackable observation likelihood and patch-onobject observation likelihood,the two likelihoods are weighted to obtain the final observation likelihood.Finally,the observation likelihood is used to calculate the weights of the particles.With the estimation result,we can predict the position and size of the target and then fulfill the object tracking task.Experiments show that our algorithm can adapt to illumination changes and large scale variations,and has a good tracking performance.2.An adaptive correlation filter based on depth features is proposed.In order to avoid the disadvantage that existing manual features cannot be applied to all general objects,pretrained VGG-Net is used for feature extraction to improve the expressiveness of features.For the purpose that adapting to partial and full occlusion,fast motion and other tracking scenes,the adaptive correlation filter is used as our base tracker,and hierarchical features learned from neural networks are used as input to track.Experimental results show that our algorithm has a good real-time tracking performance,and it is robust.3.A multi-scale correlation filter tracking algorithm is proposed.The method first obtains a multi-scale spatial image by performing Gaussian smoothing on the input image several times.Then,based on each image scale feature,a kernelized correlation filter is used to complete a tracking prediction.Finally,all prediction results are weighted,then the final position and size of the moving target are calculated.Experiments show that our algorithm can deal with complex tracking scenes and achieve accurate object tracking results.