New approaches for discovering unsupervised human activities by mining sensor data

Discovering human activities facilitates computerization and the consequent monitoring of the smart home environment. The existing unsupervised human activity discovery systems perform segmentation clustering followed by the labeling of the sensor data. Segmentation clustering consists of forming segments from similar consecutive frames and then clustering similar segments. A cluster is labeled by the action associated with its most frequent sensor(s). In these methods, even if similar segments denote distinct activities they often occur in the same cluster. We propose three alternate methods to address this issue. The first method is a minor variant of the segmentation clustering where subsequences of the segments are clustered instead of the segments. We employ the concept of cover where (a, b, c) subsumes (a, c) if they have identical frequency. The second method employs a new algorithm, i.e. LRS, instead of segmentation. The third method is a hybrid method that extracts subsequences from the output of LRS. We compared the proposed systems with the existing system on CASAS dataset, a real world human activity dataset. The third method that employs LRS followed by subsequence extraction yielded the best Dunn's index and the best correctness in clusters as per confusion matrix.