Detecting small group activities from multimodal observations

摘要

This article addresses the problem of detecting configurations and activities of small groups of people in an augmented environment. The proposed approach takes a continuous stream of observations coming from different sensors in the environment as input. The goal is to separate distinct distributions of these observations corresponding to distinct group configurations and activities. This article describes an unsupervised method based on the calculation of the Jeffrey divergence between histograms over observations. These histograms are generated from adjacent windows of variable size slid from the beginning to the end of a meeting recording. The peaks of the resulting Jeffrey divergence curves are detected using successive robust mean estimation. After a merging and filtering process, the retained peaks are used to select the best model, i.e. the best allocation of observation distributions for a meeting recording. These distinct distributions can be interpreted as distinct segments of group configuration and activity. To evaluate this approach, 5 small group meetings, one seminar and one cocktail party meeting have been recorded. The observations of the small groups meetings and the seminar were generated by a speech activity detector, while the observations of the cocktail party meeting were generated by both the speech activity detector and a visual tracking system. The authors measured the correspondence between detected segments and labeled group configurations and activities. The obtained results are promising, in particular as the method is completely unsupervised.