Trajectory mining in the context of the internet of things

© 2017 Dr. Xiaoting WangThe Internet of Things (IoT) is a technological revolution that is rapidly reshaping our society with ubiquitous sensing devices. As more sensors are being deployed, the amount of data collected is also growing significantly, leading to the increasingly important role of data mining in the IoT. One specific type of data frequently captured by sensors in the context of the IoT is trajectory data, which has attracted considerable attention lately. Various types of trajectory data have become available, such as user check-ins, vehicle GPS traces and human activity profiles recorded by smart phones. As trajectories can be generated by a large variety of sensors, many challenges arise in different application scenarios in the IoT that are associated with trajectory data mining. In this thesis, we address three trajectory data mining challenges in three different application scenarios of the IoT, namely, community, transport and healthcare applications. The first challenge is in social and community applications. The check-ins and geo-tagged photos submitted by a user on a social networking site can be regarded as trajectories that reflect the popularity of the points-of-interest (POI) being visited and the personal preference of the user. This data can be used to recommend trips to a tourist. However, the challenge of trip recommendation not only lies in searching for relevant POIs to form a personalised trip, but also selecting the best time of day to visit the POIs. Popular POIs can be too crowded during peak times, resulting in long queues and delays. To improve the quality of the solution of automated trip recommendation, we propose the Personalised Crowd-aware Trip Recommendation (PersCT) algorithm to recommend personalised trips that also avoid the most crowded times of the POIs using trajectories and pedestrian sensing data. Our results on a real-life dataset show that it is possible to achieve a balance between conflicting objectives in trip recommendation, such as satisfying user interests while reducing the crowdedness of the trips. The second challenge is in transport applications. Despite the recent introduction of advanced technologies such as Intelligent Transportation Systems (ITS), traffic congestion remains a major challenge to urban designers and transport authorities. Current ITS technologies usually employ induction loop sensors or street cameras to monitor the traffic conditions. However, monitoring the change in the traffic flows as a result of external events can be difficult. Therefore, we propose a framework to analyse changes in traffic flows due to road closure events based on GPS trajectory data. We employ ideas from contrast mining and frequent itemset mining to define, characterise and visualise the changes. The effectiveness and robustness of this framework are shown by three experiments using real taxi trajectories as well as traffic simulations in two different cities. The third challenge is in healthcare applications. The increasing health care cost has motivated the development of remote health monitoring. Home-based rehabilitation from conditions such as stroke is usually unmonitored and the progress of recovery is difficult to assess. Typically, a patient routinely revisits the hospital to seek feedback on the rehabilitation progress, which is costly and inefficient. The use of wearable sensors to capture limb movement information can be a cost-efficient alternative to frequent hospital revisits. In this context, one important problem of interest is to recognise specific actions performed by the person in daily life so that a physician can easily observe the patient’s movement from the continuous data stream sent by the sensors. We propose the use of a 3D trajectory reconstruction algorithm to process the raw sensor data and extract trajectory features that can improve action recognition accuracy. We also propose a clustering-based classifier for use with the trajectory features, and show that our proposed approach out-performs several benchmark classifiers in recognition accuracy. We conclude this thesis by presenting a number of future research directions that may be promising for trajectory mining in the IoT