This is my first ever OpenCV Haar classifier training and I need to detect people and blurr out their faces/heads/upper bodies(head+shoulders) in images and I am taking static images, where the camera does not ever move.
I started extracting positive images from sample images taken from the static camera. I started extracting heads, shoulders + head in all different orientations(back, front, sideways), since the image contains people in different positions - imagine a restaurant or a bar.
The problem is that my positive images do not contain a very distinct object but a collection of different objects ( head, shoulder + head, side head, side shoulder+head, back head, back shoulder+head). I trained the classifier with 6 phases, with only 50 positive images, and around 600 negative images(taken from an online repository). I tried the classifier to an image, and it only detected random non object pieces from the image.
I am wondering, given the background of my problem, if someone can point me in the right direction of doing this sort of detection and training - may be separating the classifier in detecting only heads, or only shoulders and head, etc... or is it okay that I mix the different positions in one classifier
I have tried all the different pregenerated classifiers that came with OpenCV and they have a very extremely low success rate on my image since the image is actually taken from higher position than the people, and it creates an angle.
