Learning-based human activity recognition

Recognizing human activities has been an extensive and interesting research topic since early 1980s. However, when deploying human activity recognition solutions to the real world, the solutions we provide must satisfy a series of requirements. We would expect our solution to be able to learn a reasonable model from as limited training data as possible. We also hope our solution would be able to deal with the complex relationships which exist in human activities. As is the case for almost all machine learning solutions, we would hope that our solution is scalable and efficient. In this thesis, we start by surveying related work and then study the solution to some specific challenges which are important to deploy these activity recognition systems in the real world.;Specifically,We first analyze how to recognize multiple activities in the physical world environment, especially when such activities have concurrent and interleaving relationships. Next, we extend such a framework to the problem of Web query classification, by exploiting the relatedness of search queries to activities with interleaving relationships and propose a context-aware query classification algorithm.;Secondly, we study the problem of abnormal activity recognition. These abnormal activities are rare to happen and it is difficult to collect enough training data about them. We design an algorithm based on the Hierarchical Dirichlet Process and the one-class Support Vector Machine to recognize abnormal activities when the training data is scarce. Finally, when we need to deploy the activity recognition systems in the real-world, it is impractical for us to collect enough training data for different activity recognition scenarios, especially when we need to collect training data for different persons and even for different actions. To solve this problem, we've developed an activity recognition framework based on transfer learning which borrows useful information from previously collected and learned activity recognition domains and then re-use such information into the new target activity recognition domain. Furthermore, we've conducted extensive experiments to demonstrate the effectiveness of our proposed approaches on real-world datasets collected from smart homes or sensor environments. We've also shown that our context-aware query classification algorithm could outperform state-of-the-art query classification approaches on real-world query engine search logs. At the end of this thesis, we discuss some possible directions and problems for future work and extensions.