Back to results list
Show full item record
Please use this identifier to cite or link to this item:
|Title:||User behaviour modelling, recognition and analytics in pervasive computing||Authors:||Liang, Guanqing||Degree:||Ph.D.||Issue Date:||2016||Abstract:||Recent years have witnessed the unprecedented growth of the adoption of sensor-rich devices such as smartphone and smartwatch, along with the large scale deployment of a variety of sensor networks in ambient environments. With the help of those ambient sensors, a large amount of users' digital traces can be collected, which opens up new opportunities for user behaviour modeling, recognition and analytics. User behaviour modeling, recognition and analytics are one of the key components in pervasive computing, which underpins a variety of significant applications: smart healthcare, business intelligence, context-aware applications, etc. Although a significant amount of research effort has been devoted in this topic, how to effectively model, recognize and analyze user behaviour still remains an open problem. In this thesis, we focus on two major research topics. First topic is how to accurately model and recognize certain categories of user behaviours based on ambient sensor data. In particular, we focus on the study of three kinds of behaviours which play critical roles in both physical and psychological heath, including: sitting posture, eating habit and social interaction. Second topic is how to identify and exploit the correlation between user behaviours and other user states such as emotion. Specifically, we conduct correlation analytics between mobility and social circle, stress and sitting posture, and exploit the correlation relationship to build up new recognition model. The details are as follows. Firstly, we aim to classify people's sitting posture based on pressure sensors-embedded seat cushion. Due to intrusiveness, high cost or low generalization accuracy, current solutions for siting posture recognition are impractical. In this work, we design Postureware, an accurate, low-cost and non-intrusive sitting posture recognition system. In particular, Postureware incorporates very thin pressure sensors to offer non-intrusive experience, an effective sensor placement solution to reduce cost, a set of user-invariant features and an ensemble learning classifier to improve generalization ability. The results show that Postureware can achieve 99.6% ten-fold cross validation accuracy and 84.7% generalization accuracy only with 10 sensors. In addition, we further evaluate the system utility by developing three applications, including unhealthy sitting posture monitoring, sitting posture-based game interface and wheelchair control. Secondly, we study the problem of recognizing people's eating behaviour using off-the-shelf smartwatch and smartphone. However, very few works have been developed for long-term eating behaviour monitoring by means of a noninvasive platform. In particular, we exploit the accelerometer of smartwatch to derive user's eating behaviour, including: eating schedule, food cuisine and food item. Besides, we leverage the collaboration between smartwatch and smartphone to reduce the energy consumption of smartwatch, and thus enabling long-term monitoring. More specifically, we propose a context-aware data collection method to conserve energy, a novel set of accelerometer features that are able to capture key characteristics of eating behaviour patterns, and a light-weight decision tree-based classification algorithm. We evaluate our approach using real-world traces and the experimental results demonstrate our work is able to monitor individual's eating behaviour in a non-invasive and energy-efficient manner.
Thirdly, we aim to model and recognize social activity based on the sensor data collected from smartphone. Most of the existing works in social activity recognition are based on the patterns of individual user such as location pattern, vocal pattern, etc. However, we observe that social activity is associated with certain group, which inherently exhibits the patterns with respect to multiple users. In this work, we introduce the concept of social circle, which reveals the behaviour pattern associated with multiple users in social activities. Here, a social circle refers to a set of users frequently gathering to conduct certain social activities. Based on the social circle concept, we present CircleSense, an accurate and efficient smartphone-based system for social activity recognition. In particular, social circle is extracted from the social proximity information obtained by Bluetooth device discovery. To further improve the system accuracy, we apply metric learning technique to extract social circle from social proximity information. To evaluate the system performance, we conduct extensive experiment based on the dataset collected in real world from 16 subjects. The experiment result shows that CircleSense outperforms the existing methods in terms of the recognition accuracy. In addition to user behaviour modelling and recognition, we study the problem of correlation analytics between user behaviours and other user states such as emotion. We find that there exists correlation between human mobility and social circle, as well as stress and sitting behaviour. By leveraging the correlation relationships, we improve the accuracy of human mobility prediction and stress measurement. In particular, we study the problem of human mobility prediction based on social context. We first conduct correlation analytics on 10-day Wi-Fi traces collected from 111K devices in a large shopping mall. We found that dwell time of repeat visitor exhibits a low degree of variation. Interestingly, visitor dwell time is positively correlated with the size of social group during the visit. By exploiting the above findings, this work presents an accurate user dwell time prediction model that incorporates time and social context, dwell time and leave time history. Evaluation results show that the proposed model is able to provide high accuracy of predicting user dwell time and outperform the baseline methods. Last but not least, we aim to measure stress based on seating pressure distribution on the chair. In particular, we collect seating pressure data from 15 participants using a seat cushion which is deployed with 20 pressure sensors. Through correlation analysis, we identify a number of seating pressure features that are associated with stress, including: average seating pressure, pressure imbalance, etc. Based on the associated features, we build up a stress detection framework to classify whether participants are stressed or not. The result show that the stress detection framework can achieve 86% accuracy using kNN classifier.
Hong Kong Polytechnic University -- Dissertations
|Pages:||xxii, 194 pages : color illustrations|
|Appears in Collections:||Thesis|
View full-text via https://theses.lib.polyu.edu.hk/handle/200/8489
Citations as of Jun 4, 2023
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.