Grounding affect recognition on a low-level description of body posture

The research presented in this thesis is centred in the rapidly growing field of affective computing and focuses on the automatic recognition of affect. Numerous diverse technologies have become part of working and social life, hence it is crucial to understand whether recognising the affective state of the user may be added to increase the technologies' effectiveness. The contributions made are the investigation of a low-level description of body posture, the proposal of a method for creating benchmarks for evaluating affective posture recognition models, and providing an understanding of how posture is used to communicate affect. Using a low-level posture description approach, this research aims to create automatic recognition models that may be easily adapted to different application contexts. These recognition models would be able to map low-level descriptions of postural configurations into discrete affective states and levels of affective dimensions. The feasibility of this approach is tested using an incremental procedure with three studies. The first study (acted postures), investigates the feasibility of recognising basic emotions and affective dimensions from acted, i.e., stereotypical, exaggerated expressions. The second study (non-acted postures), aims at recognising subtle affective states and affective dimensions from non-acted body postures in the context of a video game. In both studies, the results showed above chance level agreement and reliable consistency between human observers for the discrete affective states and valence and arousal dimensions. A feature analysis showed that specific low-level features are predictive of affect. The automatic recognition models achieved recognition rates similar to or better than the benchmarks computed. Extending the non-acted postures study, the third study focuses on how the affective posture recognition system performs when applied to sequences of non-acted static postures that have not been manually preselected, as if in a runtime situation. An automatic modelling technique was combined with a decision rule defined in this research. The results indicate that posture sequences can be recognised at well above chance level