An investigation on window size selection for human activity recognition

Blond, A., Liu, W., Cardell-Oliver, R.

Low power sensors installed in a smart house can collect data that can be mined using activity recognition techniques, allowing the actions of the resident to be inferred. This is a key step towards the dream of smart houses assisting or enhancing everyday living. Human activity recognition can be formulated as a classication problem; given a window of time, and data from sensors within that window, should it be classied as \preparing dinner", \going to bed" or \no activity"? Ecient machine learning techniques have been applied to this problem, but the discreti- sation data preparation step , also known as feature denition, selection, encoding or segmentation, despite its critical impact on the quality of the generated dataset, has received inadequate attention. In this paper, we investigate this fundamental problem of how to best discretise raw sensor data from human activities to create the feature vectors for classier training, focusing on the eect of window length on classier performance. We contribute a probabilistic model for characterising the association between window length and detection rate and introduce a modular architecture for performing experiments using different discretisation techniques. The performance of a selected Nave Bayes classier at dierent window lengths has been evaluated to demonstrate the importance of window selection.

Cite as: Blond, A., Liu, W., Cardell-Oliver, R. (2013). An investigation on window size selection for human activity recognition. In Proc. Data Mining and Analytics 2013 (AusDM'13) Canberra, Australia. CRPIT, 146. Christen, P., Kennedy, P., Liu, L., Ong, K.L., Stranieri, A. and Zhao, Y. Eds., ACS. 181-188

(from crpit.com) (local if available)