Activity Recognition on Multi-Sensor Data Streams Using Distinguishing Sequential Patterns

抄録

Activity recognition is an important task in the researches of ambient intelligence and smart environment. In real smart environments, detected signals of monitoring living space come from heterogeneous sensors generally. The intelligent awareness system can automatically provides the proper services to satisfy the requirements of users based on the recognized activity of users sensed by multi-sensor. The high accurate activity recognition is the basis of supporting high- quality service. However, the streaming data generated by multi-sensor are real time, continuous and variable. It is difficult for systems to archive high precise activity recognition from data streams of multiple sensors. The previous models on recognizing sequential data includes Hidden Markov model and Conditional Random Fields. This paper proposed a new activity recognition model for processing sequential data stream based on mining distinguishing sequential patterns. The general sequential patterns are on-line generated and counted first from the data streams and the minimal distinguishing sequential patterns are mined. Then, efficient and effective probabilistic recognizing methods and algorithms are developed for activity recognition. Two datasets, WSU and Kasteren, were used to test the proposed methods. The experimental results show that the proposed models have effective recognition rates in both of the activity level and the time slides level. The proposed model also provides a strong on-line recognition paradigm on multi-sensor data stream.