I have been implementing a visual slam approach on human skin for melanoma detection for the past few years but have reached a roadblock. I am unable to detect mathematically significant features on human skin on a 640x480 image.
I am currently using a shi-tomasi corner detector but I need to significantly reduce the corner response to detect features on parts of the human body where there does not seem to be any good features and this lets in a lot of garbage data that is in turn messing up the entire SLAM algorithm. Do you guys have any suggestions that I can use to extract data on textureless flat-regioned objects?
I have looked into most of the current implementations of feature detectors such as ORB, FAST and so on. Here is a sample image of what I am working with.
