Dan Roberts

Time series segmentation is a crucial step in the wider problem of human activity recognition. There are many algorithms that perform semantic segmentation of time series data, several of which are included in the open-source Python library aeon. How do these algorithms perform for the task of human activity recognition? Is there a better algorithm not yet included in aeon? This report implements a segmentation algorithm for time series data not currently present in the library; AutoPlait. We evaluate it and existing aeon algorithms on five human activity datasets and show that it performs in line with other algorithms currently present in aeon, achieving similar precision, recall and, F-score, in addition to providing information on any repeating or unique patterns in the data. We conclude with a critical evaluation of AutoPlait and its implementation, suggestion for further work, and a retrospective review of the project.