Research On Target Tracking Algorithm Based On Kernel Correlation Filter

Object tracking is a fundamental problem in computer vision,which supports for the high-level vision tasks such as object recognition and action recognition.Recently,although the object tracking technology has been rapidly improved,fast and accurate object tracking algorithms are still relatively scarce.In this work,we analyze the correlation filter(CF)tracking methods,and propose a novel adaptive kernel correlation filter(KCF)tracker with multiple feature integration.In addition,taking advantage of strong appearance representation of deep features,an adaptive CF tracker based on deep convolution features is proposed.Finally,since multiple object tracking(MOT)methods are widely applied in practice,we also propose an online MOT algorithm based on the adaptive CF tracker.The main innovations of this work are listed as follows:(1)In this work,an adaptive kernel correlation filter(KCF)tracker with multiple feature integration is proposed.Firstly,in order to deal with the limited appearance representation of KCF trackers,object color information is integrated into object structure features to obtain a more discriminant appearance representation.Secondly,due to the lack of scale estimation methods,KCF trackers suffer from inaccurate location estimation in case of scale variations.To deal with scale changes,we proposed a fast and accurate method to estimate the scale.In addition,since the fixed target response impedes KCF trackers performance,a novel adaptive target response method is proposed,which is calculated at each frame in the tracking process.Finally,owing to the simple linear interpolation updating method,KCF trackers are not only easy to introduce noises but also computationally expensive.We employs an adaptive high-confidence model updating algorithm.Extensive experiments on popular object tracking benchmarks show that our adaptive KCF tracker with multiple feature integration outperforms state-of-the-art tracking methods,and it achieves 79.4%and 76.0%in precision and success rate,respectively.Compare to the baseline KCF algorithm,our method obtains significant overall tracking performance improvement(11.8%in precision and 19.6%in accuracy),which can run at 32.19FPS.(2)Due to the limited discriminative ability of low-level hand-crafted features,trackers based on these features are not robust to significant appearance changes such as illumination variations,scale variations,occlusion,deformation,rotation,etc.Recently,with strong representation capabilities of deep features,we propose an adaptive KCF tracker based on deep convolutional features.We observe that the high-level features in deep neural network encode the semantic information of targets,which is more suitable for visual tracking task.However,the spatial resolution of these feature maps are too low to locate targets accurately.Therefore,we employ multiple convolution feature integration,and propose a coarse-to-fine strategy to estimate the target location.Extensive experimental evaluation on the OTB datasets show that our algorithm performs favorably against state-of-the-art methods,and it achieves 89.1%and 78.7%in precision and success rate,respectively.(3)In this work,the improved KCF tracking method is applied to MOT problem.We design two models of MOT algorithms based on the adaptive KCF tracker.At the end of this paper,a real-time on-line MOT algorithm that combines multiple target information is proposed,which utilizes the recent excellent object detection method and the Hungarian algorithm for data association.Moreover,the most of important is that our proposed MOT algorithm combines the target appearance information,position information and shape information,and it is robust to variations in tracking objects.Experimental results show that our MOT algorithm can track targets in real time accurately,and its speed can reach 25FPS in the common surveillance video.