Temporal-scale Convolutional Networks for Human Action Recognition Based on Key-Frame Extraction

Abstract

Human action recognition is an important part of intelligent video analysis. Recently, deep learning has made significant progress in this field, and state-of-the-art methods are based on the two-stream convolutional networks. In long-term action recognition, existing approaches mainly use video frames obtained by averaging or sampling as input, which may lose important information in the sampling interval. By defining the amount of video information, we propose a method of segment division and key frame extraction for action recognition is proposed, where multi-temporal-scale two-stream networks are used to extract features. We achieve 94.2% accuracy on a widely used action recognition benchmark (Split1 of UCF-101).