Online Active Learning for Human Activity Recognition from Sensory Data Streams.

Abstract

Human activity recognition (HAR) is highly relevant to many real-world do- mains like safety, security, and in particular healthcare. The current machine learning technology of HAR is highly human-dependent which makes it costly and unreliable in non-stationary environment. Existing HAR algorithms assume that training data is collected and annotated by human a prior to the training phase. Furthermore, the data is assumed to exhibit the true characteristics of the underlying distribution. In this paper, we propose a new autonomous approach that consists of novel algorithms. In particular, we adopt active learning (AL) strategy to selectively query the user/resident about the label of particular activities in order to improve the model accuracy. This strategy helps overcome the challenge of labelling sequential data with time dependency which is highly time-consuming and difficult. Because of the changes that may affect the way activities are performed, we regard sensor data as a stream and human activity learning as an online continuous process. In such process the leaner can adapt to changes, incorporate novel activities and discard obsolete ones. To this extent, we propose a novel semi-supervised classifier (OSC) that works together with a novel Bayesian stream-based active learning (BSAL). Because of the changes in the sensor layouts across different houses' settings, we use Conditional Re-stricted Boltzmann Machine (CRBM) to handle the features engineering issue by learning the features regardless of the environment settings. CRBM is then applied to extract low-level features from unlabelled raw high-dimensional activity input. The resulting approach will then tackle the challenges of activity recognition using a three-module architecture composed of a feature extractor (CRBM), an online semi-supervised classifier (OSC) equipped with BSAL. CRBM-BSAL-OSC allows completely autonomous learning that adjusts to the environment setting, explores the changes and adapt to them. The paper provides the theoretical details of the proposed approach as well as an extensive empirical study to evaluate the performance of the approach. we propose a novel semi-supervised classifier (OSC) that works together with a novel Bayesian stream-based active learning (BSAL). Because of the changes in the sensor layouts across di erent houses' settings, we use Conditional Re-