Real-time human action recognition based on person detection

Abstract

The task of human action or gesture recogntion is mostly focused on achieving good results on public datasets, but rarely achieving good results in real-time recognition. Real-time action recognition is different from training on datasets in the following aspects:1.From the saptial domain, realistic scenes are gernerally with more complicated background 2.From temporal domain, the individuals often perform actions in different lengths. In order to solve the first problem, we propose a real-time recognition method based on human body detection. We get the position of the human body in the picture through the YOLOv3 network, and then obtain the human-centered ROIs as the input of the recognition network, avoiding the interference of the background on the recognition task. Futhermore, we proposed a temporal random sampling rule that can recognize both short or long range actions. We use the indoor dangerous action detection as the application background, verify the superiority of the method on the NTURGB+D dataset, and achieve a better real-time recognition results.