A Probabilistic Framework for Detecting Unusual Events in Mobile Sensor Networks

摘要

The thesis addresses the problem of identification of unusual events in mobile sensor networks. Most existing activity recognition systems based on computer vision devices or wearable sensors attempt to recognize normal activities of daily living, however they may not be well-suited to identify abnormal activities, especially the ones that have not been encountered before. In this thesis, I will study the following research problems: 1. Learning classifiers with little or no data from unusual events. 2. Low-cost, discriminatory features and most suitable sensors to be used for unusual events detection. 3. Learning high level hierarchy of activities to help in detecting transitions to unusual events. 4. Incremental / Online learning to update the unusual events classification model dynamically as the new data is received by the sensors. To tackle these research problems, I propose to use the sequential classifiers such as Hidden Markov Models and its variants.