Research On The Algorithms Of Correlation Filter Based Visual Tracking

The objective of visual tracking is to locate a target object in continuous frames.In many intelligent applications,such as intelligent surveillance,unmanned aerial vehicle and driverless cars,visual tracking algorithm is used as one of the basic components.Thus,visual tracking is one of the foundamental problems in computer vision,which is received extensive attention.But because of the complex scenes in reality and the different,kinds of the tracking object,visual tracking becomes very challenging.There are many approaches for visual tracking problem in recent years.One of them is training the correlation filters online and detecting the position of the tracking object in every frame.Compared with other discriminative models,the correlation filter achieves good performance for visual tracking,due to its fast speed and high precision.Thus,researches on the correlation filter based tracking algorithm is very significant.In this paper,we analyze the common problems in the correlation filter based track-ers.First,the background in the sample image aggravates the boundary effects,which has negative influence to the model.Second,the tracker updates the corre-lation filter with a fixed learning rate,which corrupts the model when the object is occluded or out-of-view.Moreover,the lack of effective negative samples for correla-tion filter based tracker may lead to overfitting problem.To alleviate the problems mentioned above,we propose two trackers in this paper.Firstly,we propose a foreground weighted adaptive correlation filter based tracker.The proposed tracker can suppress the background around the target,which abates the boundary effects.Besides,to avoid the model corrupted by the occluded samples,the proposed tracker updates the models with an adaptive learning rate.We analyze a correlation filter based tracker,i.e.,STAPLE,and employ the foreground pixel scores to suppress the background in the sample image.And binarizing the pixel scores to compute the adaptive learning rate.Compared with the STAPLE tracker,the proposed tracker significantly improves the tracking performance on the OTB dataset.Besides,we propose a hard negative mining based correlation filter tracker.The proposed tracker first clusters the negative samples into several clusters,and selects hard negative samples from each cluster.To achieve better clustering results,we propose a new distance formula,which uses both the feature vectors and the spatial positions of the negative samples.Using this formula can reduce the redundancy of the negative samples and improve the training efficiency.Besides,the proposed tracker assigns an adaptive weight to each hard negative sample,which can improve the discriminative ability of the model.Compared with the STAPLE_CA tracker,which does not using hard negative samples to train the model,the proposed tracker improve the performance of the baseline tracker by a large margin.