Abstract

Self-report is the most conventional means of pain intensity assessment in clinical environments. But, it is not an accurate metric or not even possible to measure in many circumstances, e.g. intensive care units. Continuous and automatic pain level evaluation is an advantageous solution to overcome this issue. In this paper, we aim to map facial expressions to pain intensity levels. We extract well-known static (local binary pattern(LBP) and dense scale-invariant feature transform (DSIFT)) and dynamic (local binary patterns on three orthogonal planes (LBP-TOP) and three dimensional scale-invariant feature transform (3D-SIFT)) facial feature descriptors and employ the linear regression method to label a number between zero (no pain) to five (strong pain) to each testing sequence. We have evaluated our methods on the publicly available UNBC-McMaster shoulder pain expression archive database and achieved average mean square error (MSE) of 1.53 and Pearson correlation coefficient (PCC) o f 0.79 using leave-one-subject-out cross validation. Acquired results prove the superiority of dynamic facial features compared to the static ones in pain intensity determination applications.