I am working on a project which involves object detection on security recording of poor quality. I was trying to perform the detection using HOG detectMultiScale algorithm which comes with OpenCV. First I trained the algorithm using samples captured from the recording of similar quality in order to produce SVM vectors. Next I did some test but I didn't get satisfactory results. I was getting false negatives all the time.
Currently I cannot show you the actual image from the recording I am working with, but I found an exemplary image on the Internet which presents similar quality (sorry I cannot paste direct link because of insufficient points):
brightcove.vo.llnwd.net/e1/pd/4221396001/4221396001_4454349879001_4454243310001-vs.jpg?pubId=4221396001
The only difference is that my recording additionally contains some interlacing artifacts (horizontal lines), which probably makes things even worse.
I would like to ask whether it is actually possible to perform successful HOG detection on the recording of such poor quality? Note, that I don't want to detect complex shapes like faces etc, I want to detect more general, basic shapes, for example the shape of a sofa.
If HOG detection is not the way to go, could you point out some other techniques/algorithms which should give more accurate results in such use case?