Abstract
There is an increasing interest in activity recognition analysis due to the tremendous growth of sensors and devices that have recently brought significant attention to smart homes research which promotes inhabitants' comfort, security, and safety. Therefore, there is a necessity to have a comprehensive activity recognition approach which is not only limited to recognise human activities and to detect unknown/abnormal behaviors, but also could update a stream learning model in a real-time setting with sequential sensor data. As answers to the aforementioned problems, a set of approaches is developed in this thesis to fulfill these requirements with results that are comparable with or better than the state-of-the-art approaches.
By recognising activities from streaming sensor data, a new classification method, Adaptive Cluster-Based Ensemble Learning of Streaming sensor data (ACBEstreaming), is proposed. This method includes desired features, namely adaptive windowing and detecting, relevant sensor events, preserving past sensor information in its current window, and forming online clusters of streaming sensor data. ACBEstreaming improves the representation of sensor events and learns and recognises activities in a stream fashion. However, there is still the challenge of the multi-class imbalance issue in the stream learning mode that causes misleading classification outcomes due to the presence of an inadequate representation of sensor data and class distribution skews. Conversely, we propose a new multi-class stream imbalance ensemble method where the base learner is a Naïve Bayesian (NB) classifier. In this approach, the training instances from any of the classes involved in learning satisfy the median prior probability threshold to aid in balancing the classes.
Thus far, the approach is able to segment sensor data in the training part or in batch mode which requires the sensor data to have annotated labels. However, human activities have many key challenges such as their unsupervised nature in a real-time setting, online event detection, overlapping activities, and a sub-optimal choice of window size. Therefore, to address these issues, we introduce a novel real-time recognition framework which consists of Activity Features (AFs) and dynamic multi-feature windowing approaches. AFs provide statistical information about the activities from annotated sensory data in an offline phase. In the online phase, a dynamic multi-feature windowing approach using AFs and the learned NB classifier is introduced to segment unlabeled sensor data, as well as predicting the related activity, even in the presence of overlapped activities.
In addition to activity recognition analysis, we develop an Online Hidden Conditional Random Field using Resilient Gradient Algorithm (OHCRF-RGA) to address the issues of classifying sequential data where the multiple overlapping sensor-based activities have occurred. A more challenging problem is to learn unknown behaviors that have not been predefined. This is because, in a real-world environment, it is impractical to presume that users/residents will only accomplish a set of predefined activities over a long-term period. OHCRF-RGA models the sequential observations of an online stream and resolves the level of biased data.
In comparison with the-state-of-the-art approaches, we accomplished extensive experiments on five benchmark datasets acquired from residents in smart home test-beds which validate the efficiency and effectiveness of our approaches.