Action Recognition by Local Space-Time Features and Least Square Twin SVM (LS-TSVM)

摘要

In this research a new approach ffor human action recognition is proposed. At first, local spaace-time features extracted which recently becomes a popular video representation. Feature extraction is done wwith use of Harris detector algorithm and Histogram of Optiical Flow (HOF) descriptor. Then we apply a new extendedd SVM classifier called least square Twin SVM (LS-TSVM)). LS-TSVM is a binary classifier that does classification by use of two non¬parallel hyperplanes and it is four times faster than the classical SVM while the precision is better. WWe investigate the performance of LS-TSVM method on a totall of 25 persons on KTH dataset. Our experiments on the standdard KTH action dataset shown that our method improvees state-of-the-art results by achieving 95.8%, 96.3% and 97.2%% accuracy in case of 1-fold , 5-fold and 10-fold cross validation.