From conducting a thorough literature review looking at state-of-the-art affect recognition systems, it is clear that the field of affect recognition is increasing in popularity, but has many areas where further research is possible and would benefit the field.
A wide range of classification techniques have been employed in 3D facial expression recognition systems, leaving the possibility to improve on the techniques used or to combine them to produce better emotion classification results. Much research is based around the recognition and analysis of the six basic emotions, thus opportunities exist for more complex analysis that results in more diverse classifications. Recent research has shown that traditional methods of achieving facial expression recognition involving action units (AUs) and the facial action coding system (FACS) are being superseded by novel methods e.g., using muscle based features for the tracking of facial features.
Results are often provided for proto-typical lab based testing, where expressions are deliberately posed or are overly exaggerated. Real world instances of expressions are often more subtle and so systems that perform well on test data often perform significantly worse with real world data, e.g., to identify situations where there could be need for action. This is a major challenge because useful applications involve the analysis of real world situations and unconfined environments.
It is often difficult to achieve successful recognition rates with real-time performance due to the large amount of complex algorithms that have to be executed. There is often a compromise between accuracy of classification and execution speed. There is much research on facial recognition but less on body/hand and combinations of these, which provides opportunities for research focus. Giving robots emotion to allow effective human interaction is a popular research topic, with the aim of giving rise to agent-based user interfaces.
Automatic labelling systems have been developed that allow for the analysis of large datasets (in contrast to hand labelled datasets). Datasets are required for the testing of developed systems. These can prove to be problematic due to copyright restrictions limiting distribution of video clips. To overcome this, development of private datasets to evaluate and validate developed systems are commonplace. These are often small in size and limited in scope, requiring significant time to develop. Additionally, the lack of standardised datasets stops direct comparisons between research findings. Public datasets are available but there is room to develop more extensive datasets for affect recognition. It is possible to use an existing dataset to evaluate the PhD project research, or produce a custom dataset if the need is there.
Dr. Tardi Tjahjadi
BSc(Lond), MSc(UMIST), PhD(UMIST), Senior Member of IEEE
Reader (Director of Studies)
School of Engineering
University of Warwick
CV4 7AL, UK
Phone: +44 (0)24 765 23126
Fax: +44 (0)24 76 524560
T dot Tjahjadi at warwick dot ac dot uk