AbstractAutism Spectrum Disorder (ASD) is a neurodevelopmental condition that typically manifests in early childhood. Early detection and diagnosis of the disease are vital for ensuring that children receive the necessary support and interventions, which can significantly improve their outcomes. However, the shortage of clinical professionals and the lengthy diagnostic process pose significant barriers to early screening and diagnosis. This highlights the pressing need to develop efficient and effective methods to enhance the early screening and diagnosis procedure. Human behaviour analytics has emerged as a promising solution to this challenge. This field leverages computer vision technology, providing valuable insights into behaviour patterns. Utilising Vision-based human behaviour analytics methods, the atypical hand and eye behaviour patterns in individuals with ASD can be objectively measured to provide valuable insights and objective behaviour measurement to aid in the diagnosis process. This thesis aims to utilise the methods of human behaviour analytics to conduct a study of the hand and eye-related atypical behavioural patterns in individuals with ASD, with the goal of ultimately enhancing the efficiency of ASD early screening and diagnosis.
Firstly, action recognition has been incorporated into ASD research in recognising stereotyped behaviours. Compared to other modal data, skeletal data can better accommodate spatial-temporal information, which, together with the attention mechanism, is expected to model the temporal correlation better, thus improving recognition accuracy. Therefore, a self-attention network with a novel 2D skeleton joint position encoding is adopted for action/gesture recognition tasks; it obtained competitive results with reduced computational complexity. Furthermore, the action recognition performance of ASD stereotyped behaviour was significantly improved by applying transfer learning.
Secondly, to facilitate manual behaviour observation in ASD diagnoses, two behaviour quantitative analysis strategies have been proposed. The first approach proposes a Sequential Bag of Convolutional Features (SBoCF) method based on the bag-of-words (BoW) model. This approach employs deep learning techniques to eliminate the need for manually human- observed behavioural coding (HOC), enabling efficient and objective behavioural analysis. The second approach draws inspiration from the action quality assessment (AQA) concept and develops corresponding deep learning models for scoring ASD behaviours. This approach achieves promising results, showing statistical correlations between model-predicted scores and truth scores on a large public ASD dataset.
Finally, consider the atypical hand-eye behaviour in ASD to further explore joint hand-eye behaviour research and build an interpretative framework for assessing hand-eye coordination (HEC) abilities. This study designs a hand-drawing imitation protocol with a vision-based data acquisition system. A hand-eye coordination ability metric based on Cross-Recurrence Quantitative Analysis (C-RQA) is also proposed.
In conclusion, this thesis comprehensively studies hand and joint hand-eye behaviour analytics for children with ASD. By incorporating machine learning and computer vision technologies, it attempts to develop new methods and techniques to assist in the diagnosis and intervention of ASD. The study focuses on both qualitative and quantitative analysis of behaviour patterns and demonstrates the potential of these techniques in diagnosing ASD.
|Date of Award||5 Sept 2023|
|Supervisor||Honghai Liu (Supervisor) & Shikun Zhou (Supervisor)|