A Multi-resident Activity Recognition Approach based on Frequent Itemset Mining Features

Abstract

Activity recognition is one of the most important supporting technologies for smart-home applications, and most existing works conducted on this topic assume that there is only one resident in the smart home. However, there are often more than one resident at home in reality, which makes it ever more challenging to recognize their activities. In this work, a knowledge-driven approach for multi-resident activity recognition based on the features constructed by frequent itemset is proposed. Unlike traditional single-label based methods, we address this problem as learning multiple activity labels from the sensor-data sequence. Based on the binary ambient sensor data, our approach utilizes both features generated from single sensor and frequent itemset features of each class, i.e. activity. After that, five multi-label approaches are applied to recognize different activities in a short period of time. Finally, we conduct several experiments on a multi-resident activity dataset, and the experimental result demonstrates that the proposed approach with frequent itemset features is better than those without frequent itemset features added.