Vehicle verification between two nonoverlapped views using sparse representation. (September 2018)

• Record Type: Journal Article
• Authors: Hsu, Shih-Chung; Chang, I-Cheng; Huang, Chung-Lin
• Is Part Of: Pattern recognition. Volume 81(2018:Sep.)
• Page Start: 131
• Page End: 146
• Publication Date: 2018-09
• Subjects: Boost K-SVD -- K-SVD -- Vehicle verification -- Atom initializing -- Restricted isometry property (RIP); Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2018.02.031
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: The proposed system can be applied to vehicle verification under non-overlapped views of which the shapes and illuminations of vehicles are different. Propose a novel sparse dictionary learning approach, Boost K-SVD, for vehicle verification. The generated dictionary provides good RIP and sparser representation for samples. The better dictionary, the better pair verification can be promised. An adaptive dictionary size estimation is proposed to estimate optimal sizes for different datasets. Abstract: Vehicle verification in different scenes is a nontrivial problem that cannot be solved by simple correspondence matching. In the paper, the verification problem is treated as a binary classification problem. If the two vehicles in two views are the same, they are a positive pair; otherwise, a negative pair. Here, we propose an effective sparse representation (SR) method called Boost K-SVD to generate the feature vectors for vehicle representation. In Boost K-SVD, the particle filtering is first applied for the initial atom selection. Then, it finds the atoms satisfying the restricted isometry property (RIP). Finally, we propose a discrimination criterion to determine the optimal dictionary size. Instead of initial random atom selection, Boost K-SVD generates the atoms incrementally to create a more compact dictionary. Furthermore, the dictionary with RIP can produce sparser representation vectors with higher verification accuracy. Experimental results show that our method has better performance compared with the other methods.