Correlation-based real-time group activity segmentation using motion sensors

Abstract

Group Activity Recognition (GAR) is the technology learning and inferring group activity, which is done by people collaboratively, from a given data. For the real-time application of GAR, it requires a real-time activity segmentation algorithm to divide the continuously connected stream into a set of activity segments. Density ratio-based change point detection (CPD), which is recently proposed, shows promising segmentation performance on single-user activity sensor data. However, its performance on group activity data is degraded because of high false alarm rates. It is attributed to the characteristic of group activity that multiple residents act differently and simultaneously. We propose a sensor correlation-based real-time group activity segmentation methodology for detecting activity changes from concurrent sensor event stream generated by interactions among multiple users. From an active window containing active duration of motion sensors, the proposed method creates calculates previous and current correlations in each of the sensor pairs and detects changes by observing the difference of correlation between each sensor pair and centroid of sensor pairs. We evaluate our method on two group activity datasets and it shows better performance than CPD algorithms.