An end-to-end framework for the optimisation of human activity recognition

Student thesis: Doctoral Thesis


Due to recent advancements and the incessant progression of wireless sensor networks, conducting Human Activity Recognition (HAR) research within smart environments has become a widely explored domain. Nevertheless, whilst extensive research has been carried out, HAR remains a highly intricate and challenging task. Each stage of the data-driven HAR process contributes to the overall performance, thus, optimisation within each stage has driven research endeavors.

This Thesis presents an end-to-end methodology for the optimisation of HAR,
which involves investigations into enhancing performance at various key stages of the process. A publicly available HAR dataset was utilised throughout to evaluate and demonstrate the effectiveness of the proposed approach.

Initial explorations focused upon the pre-processing stage, within which the impact of data quality upon activity classification was explored using data-driven approaches to HAR. Findings demonstrated the negative impact of noise upon classification performance, with a significant performance increase of 12.97% when using cleaned data. This work led to providing recommendations as to how data should be pre-processed to prevent reductions in performance. Subsequent explorations focused upon enhancing HAR performance during the feature selection stage, within which a new hybrid feature selection method was produced. Findings revealed the effectiveness of the developed method which achieved an enhanced HAR performance of 83.24%, in addition to demonstrating the benefits of performing feature selection. A considerable trade-off was revealed between the classification performances achieved and the number of redundant features identified and removed, in comparison to the evaluated well-stablished feature selection techniques. Finally, research endeavours focused upon optimising HAR performance during the classification stage, within which both novel homogeneous and heterogeneous ensemble methods were produced. Findings demonstrated the effectiveness of the proposed ensembles, in particular the heterogeneous method which outperformed 4 benchmarked classifiers achieving an overall classification performance of 84.13%.
Date of AwardJul 2021
Original languageEnglish
SupervisorShuai Zhang (Supervisor), Hui Wang (Supervisor), Christopher Nugent (Supervisor) & Wing Ng (Supervisor)


  • Smart environments
  • Machine learning
  • Neural networks
  • Ensemble learning
  • Data-driven classification

Cite this