Temporal dependency rule learning based group activity recognition in smart spaces

Abstract

This dissertation presents a generic framework for group activity recognition using simple non-obtrusive sensors. The proposed scheme is motivated by the desire for pervasive service provision and system automation in smart spaces. Existing approaches for group activity recognition can only express the sequence of actions. Hence, they have difficulty scaling up to capture group activities. To solve this, the proposed approach is based on the idea that group activity patterns can be derived from mining interval-based temporal relationships between users' actions. It leverages a high-level hybrid architecture that can capture the essence of the temporal dependencies, represented as a set of weighed rules. The proposed approach can also learn different weights for common rules between similar group activities. To validate the proposed scheme, I used the data in Internet of Things testbed of KAIST, which was recorded over two months from a public seminar room. I compare the proposed scheme against the sequential baseline model, a mixture of Gaussian Hidden Markov Models. The proposed approach shows 1 4% better f-1 score on real data and 37% on noisy data than the baseline models. We show that the proposed model is able to distinguish between similar group activities using a small number of rules.