Conversation Activity Recognition Using Interaction Video Sequences Acquired from Groups of Pedestrians

Abstract

We propose a method for recognizing the conversation activity in a group of pedestrians walking outdoors using a camera video sequence. Various existing methods have been developed for analyzing interaction in conversation, e.g., action recognition of whether people are talking or not. However, a method has yet to be proposed to recognize the conversation activity in a group of pedestrians. In this paper, we design a method to extract features for conversation activity recognition by using body interactions that occurred from pedestrians' conversations. Our method generates an interaction video sequence in a virtual road space using a temporal pose signal and temporal position signal representing body interactions. Our method determines a conversation activity label using the interaction video sequence, which is an informative and visible feature. Experimental results show that our method using an interaction video sequence can recognize the conversation activity more accurately than comparison methods using a temporal pose signal and temporal position signal.