Datasets for Activity Recognition

We have recorded four datasets that are used in various papers in different forms. The most complete versions of the datasets can be found in the Benchmark datasets, so I recommend you use that one. However, when comparing to my work you might want to use some of the other datasets.

Benchmark datasets [link to code and datasets]

The naive Bayes, hidden Markov model, hidden semi-Markov model and conditional random field were compared for benchmark purposes in the following book chapter. Our results can be used as a baseline for comparing the performance of other pattern recognition methods (both probabilistic and non-probabilistic). The code and datasets used for the experiments can be downloaded using the link above.

Transfer Learning datasets [link to dataset]

Datasets recorded in three different houses in which the same set of activities is annotated. These datasets were used to demonstrate that the knowledge contained in activity recognition models can be transferred from one house to another.


Semi-Markov datasets [link to dataset]

A subset of datasets in which the bathroom and kitchen activities are collected in a separate dataset, based on two of the houses described above. These subsets were chosen to validate work using semi-Markov models which have a very high computational complexity. By creating subsets it was possible to validate these models on real-world data.

Ubicomp dataset [link to dataset]

First dataset recorded and used in my 2008 ubicomp paper. This dataset was the first to become publicly available and is therefore now used most in related work.