Robust Deformable and Occluded Object Tracking with Dynamic Graph
Zhaowei Cai, Longyin Wen, Zhen Lei, Nuno Vasconcelos, and Stan Z. Li

Abstract

    While some efforts have been paid to handle deformation and occlusion in visual tracking, they are still great challenges. In this paper, a dynamic graph-based tracker (DGT) is proposed to address these two challenges in a unified framework. In the dynamic target graph, nodes are the target local parts encoding appearance information, and edges are the interactions between nodes encoding inner geometric structure information. This graph representation provides much more information for tracking in the presence of deformation and occlusion. The target tracking is then formulated as tracking this dynamic undirected graph, which is also a matching problem between the target graph and the candidate graph. The local parts within the candidate graph are separated from the background with Markov random field, and spectral clustering is used to solve the graph matching. The final target state is determined through a weighted voting procedure according to the reliability of part correspondence, and refined with recourse to a foreground/background segmentation. An effective online updating mechanism is proposed to update the model, allowing DGT to robustly adapt to variations of target structure. Experimental results show improved performance over several state-of-the-art trackers, in various challenging scenarios.

Framework



Illustrative Results


Demos
    
                               avatar                                                                      basketball                                                                       bluecar                                                                        bolt


  










                               dancer                                                                         diving                                                                       football                                                                       gymnastics













                                 kwan                                                                      lemming                                                                        neymar                                                                   transformer













                               lipinski                                                                         waterski                                                                       up                                                                            yunakim













Downloads
     • Four tracking sequences collected by ourselves [download]          
     • Our tracking results for 18 sequences used in our paper [download]
     • C++ source code [DGT-v2.1]. For users in China, it could be downloaded from here [download].
    
Citations
     If you use the datasets, our tracking results or the source code, please cite our papers:
           Zhaowei Cai, Longyin Wen, Zhen Lei, Nuno Vasconcelos, and Stan Z. Li, "Robust Deformable and Occluded Object Tracking with Dynamic Graph", IEEE Transaction on Image                              Processing, 23(12), 5497-5509, December, 2014. [PDF] [BibTex]             
          •  Zhaowei Cai, Longyin Wen, Jianwei Yang, Zhen Lei, Stan Z. Li. “Structured Visual Tracking with Dynamic Graph”, 11th Asian Conference on Computer Vision, 2012. [PDF] [BibTex]       

Comments