Human action recognition using local two-stream CNN features with SVMs

Abstract

This dissertation serves to survey existing methods for human action recognition; compare those existing methods on some of the publicly-available benchmark datasets; and to introduce a novel method to solve the problem of human action recognition. The proposed method separately extracts appearance and motion features using state-of-the-art three-dimensional convolutional neural networks from sampled snippets of a video. These local features are then concatenated to form global representations for the videos. These global feature vectors are then used to train a linear SVM to perform the action classification. Additionally, we show the benefit of performing two simple, intuitive pre-processing steps, termed crop filling and optical flow scaling. We test the method extensively, and report results on the KTH and HMDB51 datasets