Automatically detecting action units from faces of pain: Comparing shape and appearance features

Abstract

Recent psychological research suggests that facial movements are a reliable measure of pain. Automatic detection of facial movements associated with pain would contribute to patient care but is technically challenging. Facial movements may be subtle and accompanied by abrupt changes in head orientation. Active appearance models (AAM) have proven robust to naturally occurring facial behavior, yet AAM-based efforts to automatically detect action units (AUs) are few. Using image data from patients with rotator-cuff injuries, we describe an AAM-based automatic system that decouples shape and appearance to detect AUs on a frame-by-frame basis. Most current approaches to AU detection use only appearance features. We explored the relative efficacy of shape and appearance for AU detection. Consistent with the experience of human observers, we found specific relationships between action units and types of facial features. Several AU (e.g. AU4, 12, and 43) were more discriminable by shape than by appearance, whilst the opposite pattern was found for others (e.g. AU6, 7 and 10). AU-specific feature sets may yield optimal results.