Multi-objective Evolutionary Computation for Human Activity Discovery

This project applies multi-objective evolutionary computation to human activity discovery: teaching an autonomous robot to find and group human activities from observation, without labelled training data. Where most human-activity work focuses on learning and recognition from known categories, this project targets the harder, less-explored discovery stage, letting the activities emerge from the data itself.

Background

Evolutionary Computation (EC) is a fast-growing area of artificial intelligence, with origins in the 1950s, that has produced human-competitive results on problems from classification and clustering to design optimisation and automatic program generation. While deep learning is widely used for computer-vision tasks such as human activity analysis, far less work has explored evolutionary techniques for these problems. Human activity analysis can be seen as a pipeline, from observing data, to discovering the activities within it, to learning a model for each, to recognising activities. Most effort goes into learning and recognition; this project concentrates on discovery.

Human activity analysis pipeline: observation, discovery, learning, recognition, with a feedback loop — The human activity analysis pipeline. This project focuses on the under-explored discovery stage, where activities are found from observation before any model is learned.

Aims and objectives

Develop multi-objective evolutionary algorithms (MOEA) for human activity discovery: build on existing unsupervised and deep-learning frameworks, test on daily-living activity data, and refine them with MOEA methods.
Develop an autonomous robot using the algorithms: find an optimal feature set for clustering human actions, let a mobile robot discover actions from its own observations, collect a substantial Activities of Daily Living (ADL) dataset, and evaluate discovery performance on public and in-house datasets.

Research team

Dr Daphne Teck Ching Lai, UBD (Principal Investigator)
Dr Ong Wee Hong, UBD (Co-Investigator)
Parham Hadikhani, UBD (PhD researcher)
Professor Yuji Sato, Hosei University, Japan (external collaborator)
Adli Mahari, UBD (Research Assistant)
Norsyamimi Razak, UBD (Research Assistant)

Publications

Lai, D. T. C., and Sato, Y. (2021). Hybrid Multiobjective Evolutionary Algorithms for Unsupervised QPSO, BBPSO and Fuzzy clustering. 2021 IEEE Congress on Evolutionary Computation (CEC), 696-703. doi:10.1109/CEC45853.2021.9504968
Hadikhani, P., Lai, D. T. C., and Ong, W. H. (2023). A Novel Skeleton-Based Human Activity Discovery Using Particle Swarm Optimization With Gaussian Mutation. IEEE Transactions on Human-Machine Systems, 53(3), 538-548. doi:10.1109/THMS.2023.3269047
Hadikhani, P., Lai, D. T. C., and Ong, W. H. (2024). Human Activity Discovery With Automatic Multi-Objective Particle Swarm Optimization Clustering With Gaussian Mutation and Game Theory. IEEE Transactions on Multimedia, 26, 420-435. doi:10.1109/TMM.2023.3266603
Hadikhani, P., Lai, D. T. C., and Ong, W. H. (2024). Flexible multi-objective particle swarm optimization clustering with game theory to address human activity discovery fully unsupervised. Image and Vision Computing, 145, 104985. doi:10.1016/j.imavis.2024.104985
Lai, D. T. C., and Hadikhani, P. (2024). Automatic Evolutionary Clustering for Human Activity Discovery. In Advances in Data Clustering (pp. 59-77). Springer. doi:10.1007/978-981-97-7679-5_4

Project details

Started 1 January 2020, duration 24 months. A previous IADA research project. Related research pages: UBD AI Lab.

Explore more on the Projects page.