Research On Collaborative Representation And Correlation Filter Based Object Tracking Algorithms

Object tracking is an important part of computer vision and its derived applications make people more convenient to enjoy their life and work.The core research of this paper is the single object tracking task.The goal of this task is to estimate the scale and location of the target in the subsequent sequences according to the information of the first frame or the previous frames.Considering the more robustness of the tracker has which can enable the tracker to deal with the influence of the diverse factors that contain occlusion,illumination variation and out of view,etc.,the more accurate predicted results it has.Therefore,how to make the tracker more robust is the goal of researchers in this field.However,due to the nature of the tracking task,which makes the information absorbed by the tracker is limited.These information parts from the annotated images before tracking and others from the previously predicted results which may contain some errors and hence influence the performance of tracker in the subsequent frames.Hence,how to promote the utilization of limited information is an important reason for promoting the robustness of the tracker.Based on the theories of collaborative representation and correlation filter,this paper presents a series of algorithms for object tracking to promote the robustness of the tracker.Specifically,the main research of this work is as follows.To alleviate the insufficient utilization of the information which is from the dictionary for the tracker,this paper presents a discriminative collaborative representation based tracking(DCRT)algorithm.The DCRT algorithm enables the proposed discriminative collaborative representation model and the state estimation model can utilize the information of the dictionary effectively.To further utilize the information of the dictionary,the DCRT algorithm brings the discriminative constraint into the discriminative collaborative representation and state estimation model.The discriminative constraint can enlarge the boundary between the positive and negative samples of the reconstruction coefficients,which can promote the discriminability and robustness of the tracker.In addition,the proposed positive sample updating scheme in the DCRT algorithm can reduce the chance of the positive sample set in the dictionary from contaminated by inaccurate tracking results.To deal with the misjudgment of the tracker which is caused by the influence of the distracters which are similar to the target,this paper proposes a distracter-aware correlation filter(DACF)for object tracking.The DACF tracking algorithm improves the regression term in the conventional correlation filter training process and searches the distracter in the response map which is acquired from the previous frame.Then,the tracker can absorb the information from the distracter and hence promotes the discriminability of the filter.Meanwhile,the DACF tracking algorithm constructs the negative filter according to the distracter-aware label map and hence presents the re-detection scheme.This scheme not only refines the tracking results but also rectifies the inaccurate tracking results from the previous frame and hence further promotes the discriminability of the tracker.In addition,the proposed scale and filter update algorithm in the proposed DACF tracking algorithm can enhance the self-adaption ability and stability of the tracker and hence makes the tracker more robust.To alleviate the imbalance of the positive and negative samples which is caused by the circular shifts strategy and the deficiency of information in negative training samples which is caused by the sine/cosine windows in the conventional correlation filter model,this paper proposes a collaborative correlation filters(CCF)for tracking.The dictionary of the CCF tracking algorithm contains three parts:historical target appearance samples,the previous tracking results samples,and the surrounding contexts samples of the former frame.However,considering the traditional correlation filter training model cannot accommodate the information of the dictionary,the CCF tracking algorithm reconstructs it and the novel model can absorb the information of the dictionary efficiently.Furthermore,the dictionary updating scheme in the CCF tracking algorithm can filter and update the predicted target,which can greatly promote the robustness of the tracker.To deal with the insufficiency of semantic information of deep features which is extracted from the training/testing samples and the inconsistency of the deep networks and the tracking model,we propose a similar-aware correlation filters network(SACFNet)for object tracking.The SACFNet tracking algorithm mainly contains off-line training and on-line tracking two parts,and each part contains the SACF model and ensemble learning algorithm.The former model can make the deep features which are extracted from the samples contains semantic information to prevent the tracker from the influence which is caused by the similar object.The later can fuse the response maps to a strong one which has more discriminability.