Object Tracking Using Minimum-Spanning-Tree One-Class Classifier

Abstract

In the traditional discriminative object tracking methods, similar to the target, background is often taken mistakenly as a single negative class to train two-class classifiers despite including various objects with different distributions. A novel method based on one-class classifier is proposed for object tracking. It takes background as a collection of all other classes except the target. Object tracking is viewed as a one-class classification problem correspondingly. Firstly, the tracking results in the recent few frames are collected as the target training samples. Secondly, a k-nearest-graph is built based on these training samples and the most representative samples are selected on the basis of their own local average degrees to construct a minimum spanning tree. Then an one-class classifier based on the minimum spanning tree is trained to model the target. Lastly the similarity scores of candidates drawn from a particle filter are evaluated by the trained one-class classifier and the one with highest score is taken as the tracking result in current frame and used to update the training sample set and the one-class classifier. Empirical results on challenging video sequences demonstrate the superior performance of our method in robustness and accuracy to state-of-the-art methods.