Adaptive updating of kernel band-width for mean-shift tracking

作者：Li, Yundong^*; Zhao, Weigang

来源：Journal of Computational Information Systems, 2012, 8(22): 9569-9579.

• Current frame
• Gaussians
• Mean shift
• Mean shift tracking
• Non-rigid objects
• Object size
• Object Tracking
• Real time tracking
• Scale parameter
• Scale-space
• Sum of gradient mode
• Target position
• Test sequence

This work presents a new approach to update kernel bandwidth in mean shift object tracking based on Gaussian scale-space. The classic mean shift tracking algorithm can't adapt to the change of object size due to its? fixed kernel bandwidth. In order to solve this problem, Gaussian scale-space is introduced to object tracking in this paper. A theorem is proved that the sum of gradient mode (SGM) of image is monotone decreasing when scale parameter is increasing in Gaussian scale-space. A novel kernel bandwidth updating method is proposed based on this theorem. Firstly target position with the most probability in the current frame is obtained through mean shift iterations, then SGMs of two successive frames are calculated. Finally, kernel bandwidth are updated according to the ratio of these two SGMs. Experiment results on various test sequences show that the proposed method can satisfy real-time tracking demand, and is effective to both rigid and non-rigid object.

• 出版日期2012
• 单位石家庄铁道大学