Research On Feature Representation And Optimization Method In Structured SVM Object Tracking

Object tracking is an important branch of computer vision,and an inter-discipline of image processing,pattern recognition,machine learning,etc.It has a broad application in areas like military guidance,visual navigation,security monitoring,human-computer interaction,and so on.Among recently proposed tracking algorithms,structured support vector machines(SSVMs)have raised widespread attention due to their novel discriminative tracking model andsuperior performance.In this work,SSVM tracking algorithm is researched from the respects of feature representation,optimization methods and anomaly detection.Theproposed methods are evaluated on object tracking benchmark(OTB).Contributions of this paper are as follows:1)An improved Struck tracking algorithm using Color Haar-like and selective updating scheme is proposed.In this work,a new feature representation method is used to incorporate color information into Haar-like features,with a gentle increase in computation.Then,a selective updating scheme is proposed to enable the Struck tracker to detect abnormal scenes and selectively stop the updating process,thus alleviating the model drift problem.Experiments on OTB50 have shown that,when utilizing Color Haar-like feature together with selective updating scheme,the Struck tracker sees increases of 9.1%,4.6%,7.5%,4.1%,8.3% and 4.5% in precision and success plots under OPE,TRE and SRE evaluations respectively.2)A dual linear SSVM tracking algorithm is proposed.In this work,a novel multi-feature target representation method is used.By combining local rank transform(LRT)feature and Lab raw feature,the details in an image are highlighted while the smooth parts are depressed.Then by using explicit feature map to discretize the multi-feature values,intersection kernel is approximated when using simply linear kernel function.From the respect of robustness enhancement,a multi-scale object detection method is used to accommodate target scale variation.Moreover,this work novelly uses dual coordinate decent(DCD)optimization method to solve the SSVM tracking model.Working together with the LRT multi-feature and explicit feature map,the proposed DLSSVM tracker performs favorably against Struck.3)A weighted margin SSVM tracking model(WMSSVM)is proposed.By using WMSSVM model,the tracker is able to take account of sample confidence at the training stage,thus adaptively learning updates from abnormal scenes.On benchmark dataset OTB100,WMSSVM algorithm achieves scores as high as 82.7%,57.5%,83.5% and 60.2% in precision and success plots under OPE and TRE evaluation metrics.In summary,three respects of SSVM tracking algorithm are studied in this paper: for feature representation,Color Haar-like and LRT multi-feature are used;for solving SSVM model,a novel DCD optimization algorithm is used;for robustness enhancement,selective updating scheme,scale estimation and weighted margin model are used.Tracking performance is improved at different levels by using the aforementioned methods.