Pain level estimation in video sequences of face using incorporation of statistical features of frames

Pain level estimation from videos of face has many benefits for clinical applications. Most of the previous works focused only on pain detection task. However, pain level estimation of video sequences has been discussed fewer. In this work, we have proposed a new regression-based approach to estimate the pain level of video sequences. As the first step, facial expression-related features were extracted from each frame, this task was done by reducing identity-related features using the robust principal component analysis decomposition. Then, we used the minimum, maximum, and mean of the features of frames in a sequence to represent that sequence by a fixed-length feature vector. After this, in order to incorporate the discriminant features of pain and reduce the computational complexity, we implemented dimension reduction by Supervised Kernel Locality Preserving Projection (SKLPP) method, and in the end, a linear regression was used to predict the pain level. Experiments on UNBC-McMaster shoulder pain expression archive database show that our method achieves the area under the curve of ROC measure of 88.43 percent in pain detection task that has better result compared to other state of the art methods.