VISUAL TRACKING VIA NONNEGATIVE MULTIPLE CODING
ABSTRACT
It has been extensively observed that an accurate appearancemodel is critical to achieving satisfactory performancefor robust object tracking. Most existing top-ranked methodsrely on linear representation over a single dictionary, whichbrings about improper understanding on the target appearance.To address this problem, in this paper, we propose a novelappearance model named as “Nonnegative Multiple Coding”(NMC) to accurately represent a target. First, a series of localdictionaries are created with different pre-defined numbers ofnearest neighbors, and then the contributions of these dictionariesare automatically learned. As a result, this ensemble ofdictionaries can comprehensively exploit the appearance informationcarried by all the constituted dictionaries. Second, theexisting methods explicitly impose the nonnegative constraintto coefficient vectors, but in the proposed model, we directlydeploy an efficient `norm regularization to achieve the similarnonnegative purpose with theoretical guarantees. Moreover, anefficient occlusion detection scheme is designed to alleviatetracking drifts, and it investigates whether negative templates areselected to represent the severely occluded target. Experimentalresults on two benchmarks demonstrate that our NMC tracker isable to achieve superior performance to state-of-the-art methods.
EXISTING SYSTEM:
Extensive research in generative methods demonstratethatsparse representation has been successfully used in visualtracking. The rationale of sparse representation based methodsis that the target can be sparsely represented by the atoms ina dictionary T (e.g. a target template set in visual trackingarea) with a sparse coefficient vector. According to the adoptedrepresentation scheme, these trackers can be grouped intoglobal template representation, local sparsemodel, joint sparse appearance model, sparse collaborative model, and structured sparsemodel. These methods learn the target representationfrom different cues and thus show promising performanceagainst various challenging factors. To be specific, in a localsparsemodel the local patches within a possibletarget candidate are sparsely represented by the patches ina dictionary. The joint sparse appearance model exploits the intrinsic relationship a particles to representthe target jointly. The sparse collaborative model combines the advantages of generative and discriminativemethods, and thus is able to exploit both holistic templates andlocal representations to describe the target. Furthermore, thestructural sparse appearance model not only exploits theintrinsic relationship a target candidates through sparserepresentations, but also retains the spatial structure a thelocal patches within every target candidate.
PROPOPOSED SYSTEM:
Based on the above observations, this paper develops a ro-bust Nonnegative Multiple Coding (NMC) tracker by exploitingan ensemble of multiple dictionaries and the nonnegativeconstraint to accurately character the target appearance.Thecontributions of this paper are summarized as follows.1) A series of local dictionaries are built with regards todifferent pre-defined numbers of nearest neighbors, andtheir related weights are automatically learned in ourNMC model. Thereby, the obtained coefficient vectorsand the ensemble of multiple dictionaries can be learnedin a uniform framework.2) We demonstrate through both theory and experimentsthat the incorporation of `norm regularization term toour NMC model is able to make the obtained coefficientvector nonnegative. It achieves the similar effect with theexact nonnegative constraint with the provable guaranteeof the lower bounded regularization parameter.23) The occlusion detection criterion is proposed to mitigatedrifting problem which investigates whether the negativetemplates (i.e. background) are used to represent thetargetV.
CONCLUSION
In this paper, we have proposed a novel NonnegativeMultiple Coding (NMC) tracker that uses an ensemble ofdictionaries to accurately character the target appearance. Inaddition, the nonnegative constraint on the coefficient vectoris replaced by an efficient `regularization term to achievethe similar nonnegative effect. Such approximation has beendemonstrated by the theoretical analysis and experimentalresults. Thereby, the obtained coefficient vector and the nonnegativeconstraint have greatly enhanced the representationability for appearance modeling of the target, which effectivelyimproves the tracking performance. The experimental resultson two benchmarks have verified the rationality of such `2norm regularization approximation and the ensemble strategy,and also demonstrated that the proposed NMC tracker achievesa favorable performance over the state-of-the-art trackers.
REFERENCES
[1] Y. Wu, J. Lim, and M. Yang, “Object tracking benchmark,” IEEE Trans.Pattern Anal. Mach. Intell., vol. 37, no. 9, pp. 1834–1848, 2015.
[2] N. Wang, J. Shi, D. Yeung, and J. Jia, “Understanding and diagnosingvisual tracking systems,” in Proc. IEEE Int. Conf. Comput. Vis. (ICCV),2015, pp. 3101–3109.
[3] Y. Yuan, H. Yang, Y. Fang, and W. Lin, “Visual object tracking bystructure complexity coefficients,” IEEE Trans. Multimedia, vol. 17, no.8, pp. 1125–1136, 2015.
[4] J. Zhang, S. Ma, and S. Sclaroff, “MEEM: Robust tracking via multipleexperts using entropy minimization,” in Proc. Eur. Conf. Comput. Vis.(ECCV), 2014, pp. 188–203.
[5] C. Gong, K. Fu, A. Loza, Q. Wu, J. Liu, and J. Yang, “PageRanktracker: From ranking to tracking,” IEEE Trans. Cybern., vol. 44, no.6, pp. 882–893, 2014.
[6] D. Wang, H. Lu, and M. Yang, “Online object tracking with sparseprototypes,” IEEE Trans. Image Process., vol. 22, no. 1, pp. 314–325,2013.
[7] J. Gao, H. Ling, W. Hu, and J. Xing, “Transfer learning based visualtracking with gaussian process regression,” in Proc. Eur. Conf. Comput.Vis. (ECCV), 2014, pp. 188–203.
[8] D. Chen, Z. Yuan, G. Hua, Y. Wu, and N. Zheng, “Descriptiondiscriminationcollaborative tracking,” in Proc. Eur. Conf. Comput. Vis.(ECCV), 2014, pp. 345–360.
[9] S. Zhang, X. Yu, Y. Sui, S. Zhao, and L. Zhang, “Object tracking withmulti-view support vector machines,” IEEE Trans. Multimedia, vol. 17,no. 3, pp. 265–278, 2015.
[10] J. Henriques, R. Caseiro, P. Martins, and J. Batista, “High-speed trackingwithkernelized correlation filters,” IEEE Trans. Pattern Anal. Mach.Intell., vol. 37, no. 3, pp. 583–596, 2015.
[11] C. Gong, D. Tao, S. Maybank, L. Wei, K. Guoliang, and Y. Jie, “Multimodalcurriculum learning for semi-supervised image classification,”IEEE Trans. Image Process., vol. 25, no. 7, pp. 3249–3260, 2016.
[12] C. Gong, D. Tao, W. Liu, L. Liu, and J. Yang, “Label propagationvia teaching-to-learn and learning-to-teach,” IEEE Trans. Neural Netw.Learn. Syst., vol. PP, no. 99, pp. 1–14, 2016.