Dynamic Near-Regular Texture Tracking

Members: Minwoo Park, Jingchen Liu, Yanxi Liu, Robert Collins

Past Memebers: Wen-Chieh Steve Lin

Topics
	Introduction Tracking Multiple Targets with Changing Topology High-Efficient Mean-shift BP for NRT Tracking Dynamic NRT Tracking and Manipulation

Introduction
	Near regular texture (NRT) is prevalent yet uncontrolled phenomenon. This implies that the analysis of NRT can give more robust solution to many real world problems. As a representative example, Lin and Liu show the dynamic NRT configuration can track dynamic NRT robustly under the challenging situations while other methods fail. Because statistical dependency of NRT components can be well encoded into graphical model, they used Markov random field to model NRT and belief propagation (BP) for inference of graph configuration. For this reason, dynamic NRT cannot go without the graphical model and, thus good inference framework in the graph is one of the fundamental components of the dynamic NRT analysis. However, the current state-of-the-art inference solution in the graph, non-parametric BP, is still slow and inaccurate for large hidden variable space. To tackle this problem, we have developed efficient and accurate mean-shift belief propagation framework. To demonstrate the efficiency, accuracy and generality, we have tested our proposed method on multi-target tracking and a 2D articulated body tracking application and show that the proposed method is better than other methods, namely BP and NBP, in terms of speed, stability and accuracy. However for the complete analysis of the Blueband, several more algorithms are needed. Another big challenge is when tracking loosely coupled textures (or multi-targets) when the topology might be changing over time. In that case, neither the target locations nor the topology is known precisely. Therefore we develop automatic spatial pattern recognition approaches and apply an EM framework by treating the multi-target spatial structure as global latent variable. And we are thus enabled to deal with multiple and changing spatial patterns.
Tracking Multiple Targets with Changing Topology
	(click on the images to download the video demos)

	[back to top]
High-Efficient Mean-shift Belief Propagation for NRT Tracking
	(click on the images to download the video demos)

	[back to top]
Dynamic NRT Tracking and Manipulation
	(click on the images to download the video demos)

	[back to top]

Publications
	Efficient Mean Shift Belief Propagation for Vision Tracking, Minwoo Park, Yanxi Liu and Robert T. Collins (CVPR '08) A Lattice-based MRF Model for Dynamic Near-Regular Texture Tracking, W.C. Lin and Y. Liu (PAMI '07) Tracking Dynamic Near-Regular Texture Under Occlusion and Rapid Movements, W.C. Lin and Y. Liu (ECCV '06)
	[back to top]