Object Tracking With Multiple-Feature Fusion

Visual target tracking is one of the most challenging research topics in the field of computer vision,and it is widely used in many fields such as intelligent monitoring,human-computer interaction,live events,and unmanned vehicle.Since the tracked object is easily affected by illumination variation,occlusion,deformation,and rotation,it is difficult to design a tracking method that is both time-effective and robust.In the past decade,a variety of different tracking algorithms have proposed to handle these challenges.Object tracking technology has also made great progress,but there are still open problems.In this dissertation,we focus on the key techniques of object tracking to solve the problems talked above using multiple feature fusion with other methods.We have proposed some robust tracking methods from the view of object appearance and discriminative tracking.The main contributions of the thesis include the following issues:In order to improve the adaptability of the apparent likelihood model based on sparse representation in a complex context,we present a novel tracking method based on collaborative sparse reconstruction of the object appearance dictionary and background dictionary.We achieve a more accurate description of the target appearance model by constructing a discriminative appearance likelihood model based on sparse representation.We embed discriminative information into the appearance likelihood model by a reasonable method of selecting the sparse coefficients of the candidate target region and the candidate background region,and by that way,we can explore the potential correlation both the candidate target region and the structure relation of the candidate background region,so as to learning the appearance model of the candidate target area more accurately.Many experimental results in challenging sequence verify the robust of our method.Our proposed tracker outperforms excellent performance in comparison with other state-of-the-art trackers.In order to handle the problem that STC adopts a single gray information feature and a simple template updating strategy and a lack of redetection mechanism make the algorithm easy to track failure caused by occlusion occurs.A multi-feature spatio-temporal context tracking method for low-rank redetection is proposed.Firstly,the multi-channel features are combined to construct the spatio-temporal context of effective compound information.This makes full use of the feature information around the target to enhance the discriminative expression of the target.Secondly,a simple and effective low-rank matrix factorization method is used to build the tracked historical information into an online redetector,which makes the tracking algorithm maintain the consistency of the structure and alleivate the object recovery problem after tracking failure.