Abstract:

In most histogram based video object tracking algorithms, an object’s histogram is typically obtained from its region in the frame where the tracking is initiated. However, single histograms are not sufficient to adapt to the object’s all possible variations. For the scenario where an object’s typical appearances are available prior to tracking, a new scale space particle filter tracking algorithm based on multiple reference histograms is proposed. We firstly employ a linear weighted method to use multiple reference histograms for producing a single histogram that is used to get the object likelihood image. Then we propose to combine the scale space theory and particle filter framework to locate objects by detecting the maximum of the multiscale normalized Laplaction filter function in the likelihood image. The comparison results with existing methods in the presented experiments on real and artificial sequences demonstrate that the proposed algorithm is able to successfully track objects whose color and brightness change drastically and rapidly, describe their size more accurately, and achieve better precision.