Robust Visual Tracking In Complicated Scenes

As one of the fundamental frontier topics in the field of computer vision,visual object tracking has numerous real applications including unmanned vehicle,intelligent surveillance and human-computer interaction.Although rapid development has been achieved by tracking algorithms in recent years,it remains challenging to achieve robust tracking due to complicated variations in appearance of objects,such as full occlusions,non-rigid deformations and rotations.In this paper,we focus on complementary feature integration,target re-detection and training samples.The major contributions of this paper are summarized as follows:To effectively improve the discriminability and robustness of the object's appearance model,we propose a complementary feature integration method.This method makes full use of the adaptability of different features,and effectively overcome the limitations caused by single feature model.Besides,the discriminability and robustness of the object's appearance model are significantly improved.For tracking failures caused by occlusions and deformations,we propose an instance specific proposals generator,which uses a prior information of the target to accurately search the target candidates.Experimental results demonstrated that our method can effectively recover the lost target caused by non-rigid deformations and occlusions.To incorporate the advantages of correlation filters and Siamese network,we treat discriminant correlation filters(DCF)as a special layer(named Correlation Filter Layer)added in Siamese network.Besides,for the problem of insufficient training samples,we propose a context-aware tracking method.This method effectively expands the number of training samples by using the context content of the target,and significantly improve the accuracy and robustness of the tracker.