Research On Robust Visual Tracking Algorithms With Distance Metric Learning

In most real tracking tasks,little object prior information is available for training,which is a great challenge for traditional visual trackers based on pre-defined distance metric.Moreover,pre-definied distance metric cannot adjust to variances in object and the environment,thus is likely to fail.As a resutl,visual trackers based on distanc emetric learning is proposed in this paper,which learn and update distance metric along the tracking procedure for improving the adaptation and accuracy of tracking.Firstly,a robust online visual tracker based on distance metric learning is proposed.The tracker regards tracking as a binary classification problem between the foreground and background information.The classifier based on Euclidean distance is updated online using information theory,thus achieving a stable tracking procedure.In order to reduce dirfts,adaptive fast updated templates and conservatively updated templates are designed and used at the same time respectively for adjusting to object variances and reduce tempalte pollution.For a faster tracking,a dense scale invariant feature transform with random principle component analysis for feature dimensionality reduction is proposed,which reduce the calculation burdon while reserving good attributes of high-dimension features.Considering object occlusion is a common challenge in visual tracking,a visual tracker based on background information is proposed.The tracker devides surroundings of the object into patches with the same size of the obejct,and decides wether occlusion occurs by comparing background patches and object patches in consequtive frames.An object detector built with random ferns is also designed for global obejct searching under severe occlusion.Scale estimation using scale pyramid is also applied for object scale variance adjustment.For evaluation of the proposed trackers,video sequences from the worldwide recognised object tracking benchmarkd are applied,which allows a testing with attributes including occlusion,deformation,rotation,scale variance,illumination changes,motion blur,etc.Center location error and overlap ratio are used as evaluation criteria.The experiment results indicate that the proposed tracker is caplable of tracking under required scenarios and is competitive comparing to state-of-art trackers.