A Knowledge-Driven Tool for Automatic Activity Dataset Annotation

Human activity recognition has become a very important research topic, due to its multiple applications in areas such as pervasive computing, surveillance, context-aware computing, ambient assistive living or social robotics. For activity recognition approaches to be properly developed and tested, annotated datasets are a key resource. However, few research works deal with activity annotation methods. In this paper, we describe a knowledge-driven approach to annotate activity datasets automatically. Minimal activity models have to be provided to the tool, which uses a novel algorithm to annotate datasets. Minimal activity models specify action patterns. Those actions are directly linked to sensor activations, which can appear in the dataset in varied orders and with interleaved actions that are not in the pattern itself. The presented algorithm finds those patterns and annotates activities accordingly. Obtained results confirm the reliability and robustness of the approach in several experiments involving noisy and changing activity executions.