Research On Wearable Sensor Network Based Human Activity Recognition Technologies

Human activity recognition technologies aim at recognizing people’s activities from the observations. As an enabling technology that makes computers capable of pro-viding active and natural services to their users, human activity recognition has many potential applications and received many research interests. In the academic world, active research has been conducted by researchers from research institutes around the world, e.g., MIT, ETH Zurich, WSU, Intel Research and Microsoft Research, etc. In the industrial world, activity recognition has been applied to systems involving mobile healthcare, mobile spot monitoring, interactive games, etc. There are two key tech-nologies that enable human activity recognition:1) activity sensing, and2) activity modeling and recognition. Traditional approach for activity sensing mainly relies on computer vision. However, the computer vision-based approaches are limited in sev-eral aspects:they are often privacy-invasive, limited in detection range and affected by many factors including light conditions and blockage. As a result, recent work proposes to use wearable sensor networks as the sensing platform and has shown its effective-ness. Regarding the problem of activity modeling and recognition, traditional work mainly focuses on recognizing a single-user’s simple activities. Most of the work can only perform off-line activity recognition. However, in real life, humans often execute activities in a complex manner, and it is often desirable to have recognition results in real-time. The above issues are not well addressed by the traditional approaches. In this thesis, we discuss the problem of complex and real-time activity recognition using efficient wearable sensor networks.To address the above problem, we first analyze the sensing and recognition prob-lems involved in a human activity recognition system, propose a general framework for a human activity recognition system, summarize and analyze the existing work relat-ed to human activity recognition. We propose three basic points to address the above problem:1) solve the problem of complex activity recognition at the activity model-ing level;2) solve the real-time activity recognition problem by designing an online and efficient system;3) address the efficient activity sensing problem by using passive sensors. We then conduct in-depth, systematical research on problems including rec-ognizing complex human activities, real-time activity recognition and efficient activity sensing. Our efforts on addressing the above issues are as follows:First, we first apply the Emerging Pattern-based approach to activity recognition. We propose a unified framework to recognize sequential, interleaving and concurren-t activities. We propose a novel pattern-Emerging Sequential Pattern, to model the sequential execution of elemental actions involved in a complex activity. We imple-ment an activity sensing platform using wearable sensor network with active sensor nodes and a prototype single user activity recognition system base on this sensing plat-form. Experiments conducted using data collected in a real smart-home environment show that the recognition accuracy of our Emerging Pattern-and Emerging Sequen-tial Pattern-based approaches have outperform the traditional approaches, which are88.11%and91.89%, respectively.Second, we are among the first to discuss the multi-user activity recognition prob-lem in the area of wearable sensor networks. We propose to use two graphic models, i.e., Coupled Hidden Markov Model and Factorial Conditional Random Field to model and recognize multi-user activities. Base on the Emerging Pattern-based single user activity, we propose an Emerging Pattern-based multi-user activity model to recog-nize multi-user activities. We implement a prototype multi-user activity recognition system base on the wearable sensor network-based sensing platform and conducted extensive experiments to evaluate the proposed apporaches. The results show that our Coupled Hidden Markov Model-, Factorial Conditional Random Field-, and Emerging Pattern-based approaches have achieved an accuracy of85.46%,86.54%, and89.72%, respectively.Third, regarding the real-time activity recognition problem, we discuss two types of real-time requirements-soft real-time and hard real-time. Regarding the soft real-time activity recognition problem, we propose a sensor-based gesture recognition al-gorithm and an Emerging Pattern-based online, fast activity recognition algorithm to perform soft real-time activity recognition. We implement the prototype system and e- valuate it using the single user activity data set. The experiment results show we reduce the communication cost by60.2%and achieve a recognition accuracy of82.87%with an average delay of5.7s. Regarding the hard real-time activity recognition problem, we first propose two requirements, i.e., online and continuous, for an activity recognition system to achiever hard real-time. We then propose to use fixed sliding-window-based data segmentation technology and a Support Vector Machine-based fast recognition al-gorithm to achieve these two requirements. We implemented a prototype of the hard real-time activity recognition system base on a passive sensing platform and evaluated it with data collected in a real environment. The results show our system can perform real-time recognition even if the delay bound is1s and when the delay bound is5s, our system achieves the best recognition accuracy of93.6%.Finally, to achieve efficient activity sensing, we design and implement a nov-el, passive UHF RFID-based activity sensing platform which is more cost-effective than the previous active sensor-based activity sensing platform. We propose two al-gorithms:the data completion algorithm and feature extraction algorithm to address the challenges related to the RFID technology. The passive sensing platform is used to implement the prototype hard real-time activity recognition system and achieved a recognition accuracy of93.6%in our experiments, which demonstrates the effective-ness and efficiency of using the passive sensing platform for activity recognition.