Recognition of interactive human groups from mobile sensing data

In real life, people often participate in activities in groups. During the activities, group members commonly engage in interactions such as shaking hands, waving hands, embracing, and hooking arms. Existing approaches to recognize human groups assume that the individuals’ locations or sensing signals are similar; the interactions among them are probably regarded as dissimilar data and affect the recognition accuracy. Moreover, not all persons undertake the same interactions simultaneously. In this study, we propose an approach named interactive group recognizing (IGR) to solve this problem. We collected the sensing data from individuals to deduce their interactions, and compute the disparity between two individuals. Subsequently, a majority-voting based method is applied to recognize the human groups to eliminate the inconsistency among interactions. We also analyzed the number of interactions that could occur without adversely affecting the performance of our approach. Compared with the results from existing approaches, divergence-based affiliation detection (DBAD) and cross-correlation based approach, IGR improved the group recognition accuracy by 6.9% and 26.7%, respectively, and F1-score by 13.6% and 54.6%, respectively, when interactions among people account for no less than 8% of the total execution time.