Group Activity Recognition Using Wearable Sensing Devices

Understanding the behavior of groups or crowds in real time can provide valuable information to crowd management systems, helping prevent or avoid human tragedy in crowd emergencies. Wearable devices provide a powerful platform for understanding human behavior, however the infrastructure required to communicate data from these devices is often the first casualty in emergency situations. Peer-to-peer (P2P) methods for recognizing group behavior are therefore necessary, but the behavior of the group cannot be observed at any single location, creating an intriguing problem. This dissertation provides the tools to (1) understand which information is best for behavior recognition, (2) to detect different groups who may be in the same environment, and (3) to recognize the physical behavior or activities of the group, all in a P2P fashion. Furthermore, all of this is done while (4) respecting the limited resources and primary functions of the sensing devices, e.g. wearables and mobile phones. The combined contribution of this dissertation is the knowledge, algorithms and methods necessary for recognition of group behavior using only the wearable devices of its constituents.