Research On Ensemble Learning Based Object Tracking Algorithms

The technology of visual tracking found a wide utilization in fields such as security monitoring and intelligent transportation due to its strong practicability.However,it remains challenging to achieve a robust and accurate tracking of the targetsome due to the known factors,such as the influence of the deformation of the target or illumination changing.To find a tacking algorithm adaptive to complex surrounding environment,many researchers find their way out by employing learning models with the help of Ensemble Learning,in order to obtain a tracker of superior performance and higher earning capacity compared with the original weak learning models.In the recent years,many algorithms attempt to apply discriminative classifiers to target tracking,which transform target recognition into target-background classification.These algorithms acquire the change of target appearance by online learning,and then adapt to the change of target advancing with time,so that the tracking process can be completed robustly in complex motion scene.Among them,there are algorithms acheive the mentioned goal by simply utilizing the ensemble learning framework which makes full use of the advantages of the base classifiers to construct strong classifier.In this paper,we made a research on the two commonly used algorithms as known as SVM(Support Vector Machine)and KCF(Kernelized Correlation Filter).We had also made a research on the cases of ensembling SVM classifiers and KCF filters respectively,and 2 novel improved algorithms are proposed respectively.To summarize,we make the following contributions in this paper:(1)A tracking algorithm based on Boosting ensembled SVMs is proposed.The base classifier pool is built by combining different features and SVM kernel functions,which is the source of the difference of the SVM based classifier.During the process of combining the base classifiers into a strong classifier,the base classifiers are trained by different kernel functions and different features,while the algorithm in earlier years would only train the base classifiers on either difference kernel functions or features.Comparative tests with single-kernel-SVM tracker also demonstrate that Ensemble Learning makes the base classifiers improved.(2)A tracking algorithm based on ensembled KCFs is proposed to overcome the shortback of a restriction to scale estimation of the original KCFs by adding an independent scale detector.To reduce computational cost,the scale detector is trained on the features compressed in the Fourier domain.In addition,the original KCF tracker is trained only on gray HoG features,while the crucial color characteristics of the target information is missing,so the proposed algorithm adds another KCF tracker trained on Color Name features during the process of the tracker generation to make up for the inadequacy in target appearance model.Moreover,in order to improve the accuracy of the algorithm,the 2 trackers are ensembled under the cooperative-training framework.The experiment result show that the proposed algorithm has higher tracking accuracy than the original KCF tracker,and is also adaptive to the scale change of the target.