Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.705753
Title: Machine learning for automatic analysis of affective behaviour
Author: Nicolaou, Michael (Mihalis)
ISNI:       0000 0004 6061 4207
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2014
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Abstract:
The automated analysis of affect has been gaining rapidly increasing attention by researchers over the past two decades, as it constitutes a fundamental step towards achieving next-generation computing technologies and integrating them into everyday life (e.g. via affect-aware, user-adaptive interfaces, medical imaging, health assessment, ambient intelligence etc.). The work presented in this thesis focuses on several fundamental problems manifesting in the course towards the achievement of reliable, accurate and robust affect sensing systems. In more detail, the motivation behind this work lies in recent developments in the field, namely (i) the creation of large, audiovisual databases for affect analysis in the so-called ''Big-Data'' era, along with (ii) the need to deploy systems under demanding, real-world conditions. These developments led to the requirement for the analysis of emotion expressions continuously in time, instead of merely processing static images, thus unveiling the wide range of temporal dynamics related to human behaviour to researchers. The latter entails another deviation from the traditional line of research in the field: instead of focusing on predicting posed, discrete basic emotions (happiness, surprise etc.), it became necessary to focus on spontaneous, naturalistic expressions captured under settings more proximal to real-world conditions, utilising more expressive emotion descriptions than a set of discrete labels. To this end, the main motivation of this thesis is to deal with challenges arising from the adoption of continuous dimensional emotion descriptions under naturalistic scenarios, considered to capture a much wider spectrum of expressive variability than basic emotions, and most importantly model emotional states which are commonly expressed by humans in their everyday life. In the first part of this thesis, we attempt to demystify the quite unexplored problem of predicting continuous emotional dimensions. This work is amongst the first to explore the problem of predicting emotion dimensions via multi-modal fusion, utilising facial expressions, auditory cues and shoulder gestures. A major contribution of the work presented in this thesis lies in proposing the utilisation of various relationships exhibited by emotion dimensions in order to improve the prediction accuracy of machine learning methods - an idea which has been taken on by other researchers in the field since. In order to experimentally evaluate this, we extend methods such as the Long Short-Term Memory Neural Networks (LSTM), the Relevance Vector Machine (RVM) and Canonical Correlation Analysis (CCA) in order to exploit output relationships in learning. As it is shown, this increases the accuracy of machine learning models applied to this task. The annotation of continuous dimensional emotions is a tedious task, highly prone to the influence of various types of noise. Performed real-time by several annotators (usually experts), the annotation process can be heavily biased by factors such as subjective interpretations of the emotional states observed, the inherent ambiguity of labels related to human behaviour, the varying reaction lags exhibited by each annotator as well as other factors such as input device noise and annotation errors. In effect, the annotations manifest a strong spatio-temporal annotator-specific bias. Failing to properly deal with annotation bias and noise leads to an inaccurate ground truth, and therefore to ill-generalisable machine learning models. This deems the proper fusion of multiple annotations, and the inference of a clean, corrected version of the ''ground truth'' as one of the most significant challenges in the area. A highly important contribution of this thesis lies in the introduction of Dynamic Probabilistic Canonical Correlation Analysis (DPCCA), a method aimed at fusing noisy continuous annotations. By adopting a private-shared space model, we isolate the individual characteristics that are annotator-specific and not shared, while most importantly we model the common, underlying annotation which is shared by annotators (i.e., the derived ground truth). By further learning temporal dynamics and incorporating a time-warping process, we are able to derive a clean version of the ground truth given multiple annotations, eliminating temporal discrepancies and other nuisances. The integration of the temporal alignment process within the proposed private-shared space model deems DPCCA suitable for the problem of temporally aligning human behaviour; that is, given temporally unsynchronised sequences (e.g., videos of two persons smiling), the goal is to generate the temporally synchronised sequences (e.g., the smile apex should co-occur in the videos). Temporal alignment is an important problem for many applications where multiple datasets need to be aligned in time. Furthermore, it is particularly suitable for the analysis of facial expressions, where the activation of facial muscles (Action Units) typically follows a set of predefined temporal phases. A highly challenging scenario is when the observations are perturbed by gross, non-Gaussian noise (e.g., occlusions), as is often the case when analysing data acquired under real-world conditions. To account for non-Gaussian noise, a robust variant of Canonical Correlation Analysis (RCCA) for robust fusion and temporal alignment is proposed. The model captures the shared, low-rank subspace of the observations, isolating the gross noise in a sparse noise term. RCCA is amongst the first robust variants of CCA proposed in literature, and as we show in related experiments outperforms other, state-of-the-art methods for related tasks such as the fusion of multiple modalities under gross noise. Beyond private-shared space models, Component Analysis (CA) is an integral component of most computer vision systems, particularly in terms of reducing the usually high-dimensional input spaces in a meaningful manner pertaining to the task-at-hand (e.g., prediction, clustering). A final, significant contribution of this thesis lies in proposing the first unifying framework for probabilistic component analysis. The proposed framework covers most well-known CA methods, such as Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), Locality Preserving Projections (LPP) and Slow Feature Analysis (SFA), providing further theoretical insights into the workings of CA. Moreover, the proposed framework is highly flexible, enabling novel CA methods to be generated by simply manipulating the connectivity of latent variables (i.e. the latent neighbourhood). As shown experimentally, methods derived via the proposed framework outperform other equivalents in several problems related to affect sensing and facial expression analysis, while providing advantages such as reduced complexity and explicit variance modelling.
Supervisor: Maja, Pantic Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.705753  DOI: Not available
Share: