期刊名称：International Journal of Multimedia and Ubiquitous Engineering
印刷版ISSN：1975-0080
出版年度：2015
卷号：10
期号：1
页码：391-402
DOI：10.14257/ijmue.2015.10.1.36
出版社：SERSC
摘要：To utilize the geometry structure information and similarity information within the neighborhood surrounding a spatio-temporal interest point for human action recognition task, we employ the axes of a regular polyhedron as a reference locating system, and build a novel local feature named polyhedron neighborhood feature (PNF). Then, to reduce quantization error in the coding stage, locality-constrained linear coding method is used to encode the obtained PNFs. Next, multi-temporal-scale PNFs (MPNFs) are created for handling the problem of various action speeds. In classification, support vector machine (SVM) based on linear kernel is used as classifier taking time consumption into account. The experiments on the KTH and UCF sports datasets show that the recognition system based on PNFs achieves better performance than the competing local spatio-temporal feature-based human action recognition methods.
