Human activity recognition: classifier performance evaluation on multiple datasets
Pavel Dohnálek, Petr Gajdoš, Tomáš Peterek · Journal of Vibroengineering · 2014
Human activity recognition is an active research area with new datasets and new methods of solving the problem emerging every year. In this paper, we focus on evaluating the performance of both classic and less commonly known classifiers with application to three distinct human activity recognition datasets freely available in the UCI Machine Learning Repository. During the research, we placed considerable limitations on how to approach the problem. We decided to test the classifiers on raw, unprocessed data received directly from the sensors and attempt to classify it in every single time-point, thus ignoring potentially beneficial properties of the provided time-series. This approach is beneficial as it alleviates the problem of classifiers having to be fast enough to process data coming from the sensors in real-time. The results show that even under these heavy restrictions, it is possible to achieve classification accuracy of up to 98.16 %. Implicitly, the results also suggest which of the three sensor configurations is the most suitable for this particular setting of the human activity recognition problem.