A rule-based approach to detect groups using a depth sensing camera

Abstract

In this work we present a rule-based group detection strategy. To achieve this we created a library with an event-driven architecture that facilitates the development of proxemic and group-aware applications. Our method to detect and track groups in a crowd uses a rule set based on basic metrics that are calculated using the frames captured by a Microsoft Kinect sensor. This rule set consists of 7 rules. The specific rules are minimum inter- personal distance, minimum difference in average speed, minimum difference in distance-to-screen, difference in initial and current direction, difference in the arrival times and the time all rules are met. To find the best values of the parameters we conducted observational tests of 9 different situations. We found that the best rule set is highly dependent on the situation and varies across different applications. We describe various research studies concerning group detection and tracking to provide a context of the research domain before we describe the rule set and the library we developed which enables developers to tune this rule set for any specific application. As an example we created a basic group-aware game that serves as a demo of how to use our library. We also note observations with regard to this game in this work.