RGBT Object Tracking Based On Correlation Filter

RGBT target tracking uses visible light(RGB)and thermal infrared(T)data to achieve robust target tracking in different fusion ways.Single modality(RGB)target tracking vulnerable to illumination intensity,occlusion and weather environment,while thermal infrared technology is not affected by the above factors due to its unique thermal sensitivity during target tracking.Therefore,RGBT target tracking in video surveillance,security,object detection,and other areas of the application of more scientific significance and application value.However,RGBT target tracking algorithm based on correlation-filtering model is still facing some challenges,many RGBT target tracking algorithm is difficult to implement modal adaptive fusion and target tracking method based on correlation-filtering mostly directly use feature map to train samples,it may introduce noise and redundancy,thus reducing the robustness of tracking.To solve the above problems,this dissertation study RGBT target tracking under the framework of correlation-filtering,to achieve a robust tracking effect and real-time tracking speed.This dissertation proposes the following two visual collaboration models:First,a collaborative correlation-filtering model is proposed.Many current RGBT target tracking algorithms are difficult to achieve real-time tracking speed,and most RGBT tracking algorithms ignore the modal quality perception assessment,which easily affects tracking performance.To solve the above problems,a RGBT target tracking method based on collaborative correlation model is proposed.First,the adaptive weight allocation strategy is adopted to evaluate the quality of the two modalities,and the advantages and disadvantages of the two modalities are obtained by comparing them,and then different weight values are assigned.Second,considering that the target is blocked for a long time or the target is beyond the search area will lead to the loss of the target,which will affect the tracking results,the dissertation uses re-detection technology to judge whether the target is lost and prevent the tracking failure.Simultaneously,an efficient alternative direction multiplier method is used to optimize the collaborative correlation filter to improve the tracking speed.Finally,experimental results on GTOT and RGBT210 datasets show that the proposed method achieves effective information fusion between modalities,and achieves a robust tracking effect and real-time tracking speed.Second,the feature-coding correlation filter model is proposed.The target tracking algorithm based on the correlation filter model adopts the pre-trained convolutional neural network model,and the depth features obtained are usually redundant and noisy.In addition,the first work did not consider the processing problem of target features,nor did it fully mine the feature information,which affected the tracking effect.To solve this problem,an RGBT target tracking algorithm based on feature-coding correlation-filtering is proposed.Specifically,the method has good discriminability and generalization.Under the framework of discriminant correlation filter,the dependency between local features is used to encode the features of the two modalities,and the noise and redundant information are effectively reduced.Then,Laplacian-coding algorithms are used to learn compact,differentiated,and object-oriented feature representations.Simultaneously,a fast solution method is adopted to reduce the tracking computation,so that the feature representation and the correlation filter are mutually enhanced.Finally,a large number of experimental results show that the tracking performance of this method is better than that of the benchmark tracker,and it has little effect on the tracking speed.