Inferring Facial and Body Language

Abstract

Machine analysis of human facial and body language is a challenging topic in computer vision, impacting on important applications such as human-computer interaction and visual surveillance. In this thesis, we present research building towards computational frameworks capable of automatically understanding facial expression and behavioural body language. The thesis work commences with a thorough examination in issues surrounding facial representation based on Local Binary Patterns (LBP). Extensive experiments with different machine learning techniques demonstrate that LBP features are efficient and effective for person-independent facial expression recognition, even in low-resolution settings. We then present and evaluate a conditional mutual information based algorithm to efficiently learn the most discriminative LBP features, and show the best recognition performance is obtained by using SVM classifiers with the selected LBP features. However, the recognition is performed on static images without exploiting temporal behaviors of facial expression. Subsequently we present a method to capture and represent temporal dynamics of facial expression by discovering the underlying low-dimensional manifold. Locality Preserving Projections (LPP) is exploited to learn the expression manifold in the LBP based appearance feature space. By deriving a universal discriminant expression subspace using a supervised LPP, we can effectively align manifolds of different subjects on a generalised expression manifold. Different linear subspace methods are comprehensively evaluated in expression subspace learning. We formulate and evaluate a Bayesian framework for dynamic facial expression recognition employing the derived manifold representation. However, the manifold representation only addresses temporal correlations of the whole face image, does not consider spatial-temporal correlations among different facial regions. We then employ Canonical Correlation Analysis (CCA) to capture correlations among face parts. To overcome the inherent limitations of classical CCA for image data, we introduce and formalise a novel Matrix-based CCA (MCCA), which can better measure correlations in 2D image data. We show this technique can provide superior performance in regression and recognition tasks, whilst requiring significantly fewer canonical factors. All the above work focuses on facial expressions. However, the face is usually perceived not as an isolated object but as an integrated part of the whole body, and the visual channel combining facial and bodily expressions is most informative. Finally we investigate two understudied problems in body language analysis, gait-based gender discrimination and affective body gesture recognition. To effectively combine face and body cues, CCA is adopted to establish the relationship between the two modalities, and derive a semantic joint feature space for the feature-level fusion. Experiments on large data sets demonstrate that our multimodal systems achieve the superior performance in gender discrimination and affective state analysis.