Multimodal analysis of free-standing conversational groups

Abstract

"Free-standing conversational groups" are what we call the elementary building blocks of social interactions formed in settings when people are standing and congregate in groups. The automatic detection, analysis, and tracking of such structural conversational units captured on camera poses many interesting challenges for the research community. First, although delineating these formations is strongly linked to other behavioral cues such as head and body poses, finding methods that successfully describe and exploit these links is not obvious. Second, the use of visual data is crucial, but when analyzing crowded scenes, one must account for occlusions and low-resolution images. In this regard, the use of other sensing technologies such as wearable devices can facilitate the analysis of social interactions by complementing the visual information. Yet the exploitation of multiple modalities poses other challenges in terms of data synchronization, calibration, and fusion. In this chapter, we discuss recent advances in multimodal social scene analysis, in particular for the detection of conversational groups or F-formations [Kendon 1990]. More precisely, a multimodal joint head and body pose estimator is described and compared to other recent approaches for head and body pose estimation and F-formation detection. Experimental results on the recently published SALSA dataset are reported, they evidence the long road toward a fully automated high-precision social scene analysis framework.