Human Activity Recognition and Behaviour Understanding


Introduction

This project investigates computational methods for recognizing human actions, interpreting body movements, and identifying behavioural patterns from visual and sensor data. This work spans skeleton-based action recognition [Men et al. Neurocomputing2023, Ho et al. VRST2015], smartphone-based activity recognition using inertial sensor signals [Thakur et al. IRI2023, Thakur et al. Access2022], depth-based posture classification [Ho et al. CVIU2016], and unsupervised abnormal behaviour detection in crowded environments [Xu et al. SKIMA2017]. By modelling spatial relationships between body parts and learning temporal patterns in movement, we aim to develop robust methods for understanding human activities across different environments and sensing modalities.

A key direction of this research is developing effective representations of human movement that generalise across actions, viewpoints, and data conditions [Ho et al. CVIU2016]. Our work explores self-supervised contrastive learning for view-invariant skeleton-based action recognition [Men et al. Neurocomputing2023], alongside deep architectures that combine convolutional autoencoders and recurrent networks to learn discriminative temporal features from smartphone sensor data [Thakur et al. IRI2023, Thakur et al. Access2022]. Together, these studies advance machine learning approaches for human activity recognition and behaviour analysis, with potential applications in smart environments, activity monitoring, and automated behavioural understanding.

Publications


The Team