Research Of Correlation Filter Tracking With The Integration Of Convolutional Neural Network

General visual object tracking is a very important and active area of research in computer vision,and it has a wide range of applications in video surveillance,human-computer interaction and autonomous driving.For a long time,domestic and foreign researchers have proposed a number of solutions towards object tracking.Among them,the method based on correlation filter is known for its high efficiency,and has received more and more attention in recent years.Most of the current works have promoted the accuracy performance by improving the correlation filter algorithm,but it has greatly reduced the efficiency of the algorithm.This paper deeply analyzes and reveals the shortcomings of the correlation filter tracking algorithm,and proposes to effectively solve these problems with the integration of convolutional neural network,and finally achieves good and fast object tracking algorithms.We implement two correlation filter tracking algorithms with the integration of convolutional neural network.The main works include:First,the correlation filter tracker with siamese network.Firstly,in the correlation filter tracking algorithm,filters need to be updated online to adapt to the dynamic changes of the target appearance in the video.At present,most algorithms directly use the tracking result of each frame for filters updating,so when the target is severely occluded,the updated filters will over-fit the background.Aiming at this problem,this paper implements a robust and adaptive filters updating strategy by using the siamese network to evaluate the tracking results.Then,although the classical correlation filter trackers have high efficiency,they can only detect the target in a limited search region due to the influence of boundary effect.In addition,due to the use of hand-crafted features,these algorithms often fail to track the target while encountering fast target motion and background clutters.In response to these shortcomings,we propose a multimodal detection strategy.It can effectively reduce the tracking failure caused by the small search area and weak feature representation by adaptively generating candidate targets and verifying them with siamese network.Finally,we propose to build a template library online during the object tracking process.A diverse library of templates allows the siamese network to have stronger discriminative ability than using only one fixed template.Second,background-aware correlation filter network.The correlation filter tracking framework mainly includes two parts: feature representation and filter learning.In terms of feature representation,hand-crafted feature or convolutional feature from other task are used by most correlation filter algorithms.We consider the joint learning of feature representation and filter learning in order to learn the proper feature representation for filters.Specifically,we construct a convolutional neural network and interpret the filters as a layer with differentiable properties in the network,called the correlation filter layer,and then we can end-to-end train the network through backpropagation algorithm.As a result,we can learn the proper feature representation.Compared to the use of classical correlation filter tracker as the correlation filter layer,we propose to use a more advanced one,and finally obtain a background-aware correlation filter network for object tracking.A large number of comparative experiments on the datasets OTB-13 and OTB-15 show that the two proposed correlation filter tracking algorithms combined with convolutional neural networks achieve superior tracking accuracy while maintaining real-time tracking performance.On the OTB-15,the correlation filter tracker with siamese network algorithm achieves 61.3% AUC score,which improves the baseline to a relative gain of 12.3%.The background-aware correlation filter network achieves 62.6% AUC score,which improves the baseline to a relative gain of 11.0%.