Congratulations to Amran! His paper “Investigation of the Unsupervised Machine Learning Techniques for Human Activity Discovery” received the Gold Award in General Electronic Category at the second International Conference on Electronics, Biomedical Engineering and Health Informatics (ICEBEHI) 2021. Amran presented the paper virtually held on 3-4 November 2021.
Abstract. Human activity recognition has been considered as the main capability of an intelligent system in understanding of human activities. Human activity recognition focuses on classifying activities with predefined models learned from labelled data based on supervised or semi-supervised approaches. These ap-proaches have assumed the availability of abundant labelled activity observations. In real-world scenarios, labelled activity observations are difficult to obtain given the undefined number of human activities and their wide variation between differ-ent subjects. The desirable approach is an un-supervised one in which an intelli-gent system can discover new activities from unlabeled observations. This work aimed to evaluate the performance of several clustering algorithms to effectively distinguish different daily activities for human activity discovery. Clustering algo-rithms used include k-means, spectral, hierarchical and BIRCH clustering. Activi-ty observations were represented as a sequence of postures with 3D skeletal joint locations derived from the Kinect depth map, and then different clustering algo-rithms were applied to the data. The approach is evaluated on a lab recorded da-taset and a publicly available dataset. Overall mean precision, recall and F1-score for both datasets were above 58%, 68%, 61% respectively. K-means and ag-glomerative clustering with ward linkage achieved highest precision, recall and f1-score on both datasets which demonstrated the potential of using clustering al-gorithms to distinguish and group different activities for activity discovery with-out using labeled data.
More info on the project can be found here.