(Re)moving Parts: Towards a System for the Separation of Affective Movements in Video Data

In conversation, humans make facial movements for a Variety of reasons. Some movements serve a primarily functional role, such as blinks that moisten the eyes. Other movements support a different goal, such as the movements of lips in speech. Still others are affective movements which display an emotional state of the actor, whether accurately or inaccurately. Each of these streams of movement evolves at a different timescale. As a result, it should be possible to separate them based on dynamic frequency information. Facial movement data can be extracted from Video using computer Vision techniques, but no tool exists for separating different streams of movement from these combined data. This dissertation proposes several methods for separating these streams, including a novel wavelet-based method. These methods are tested against each other using a set of simulated signals built to resemble the hypothesized mixture of source streams. Results indicate that dynamics - related methods such as Slow Feature Analysis may be suitable tools for separating such streams, but further study will be required to determine the best method.

Note: Abstract extracted from PDF file via OCR