EK-means tracker: A pixel-wise tracking algorithm using kinect

Abstract

This paper describes a novel object-tracking algorithm by classifying the pixels in a search area into “target” and “background” with K-means clustering algorithm. Two improvements are made to the conventional K-means tracker to solve the instability problem that occurs when some background objects show similar colors to the target or the size of the target object changes significantly. The first one is introducing of the depth information as the sixth feature into the original 5D feature space for describing pixels. The second one is to use Mahalanobis distance in order to keep the balance between color and position when evaluating the difference between pixels. EK-means Tracker can track non-rigid object and wired object at video rate. Its effectiveness was confirmed through several comparison experiments.