Hierarchical Least Square Twin Support Vector Machines Based Framework for Human Action Recognition

The aim of this paper is presentation of a new human action recognition framework. In the proposed framework, local space-time features extracted by use of Harris detector algorithm and Histogram of Optical Flow (HOF). A new classifier based on two non-parallel hyperplanes called Twin Support Vector Machines (TWSVM) is used which is four times faster than classical SVM. According to the prior knowledge that two classes of human action recognition (jogging and running) are very similar and recognition of these classes are difficult, a hierarchical structure is used for better recognition. We applied our method to KTH dataset to investigate the performance of the proposed action recognition approach. Our experimental result shown that our approach improves state-of-the-art results by achieving 98.33%, 96.39% in case of leave-one-out and 10-fold cross validation.