I am trying to deal with the anomaly action detection task and the first step is to obtain skeleton information using pose detectors. According to the paper that I refer to I need seven number, the two-dimensional joint coordinates on the image, the time index, the joint confidence, the joint index, and the two-dimensional centroid coordinates calculated from the human joints. I am quite unsure how to calculate the centroid. Does it mean that I need to calculate the average of the joint coordinates for each person which the pose detector detects, even some of the joints aren't detected?
I tried to get skeleton information in some videos using OpenPose, while the result seemed to be unsatisfied, which makes me more worry about my expected method to calculate centroid coordinates.