combining two human detection methods
1 view (last 30 days)
I am trying to detect people in side a home, I tried HOG once and Viola-object detection upper body another time.
now I am thinking to combine between these two methods, could you please show me the ways for combining as I so junior in image processing.
thanks very much for any cooperation
Dima Lisin on 15 Feb 2016
To combine these two methods, you first have to run the image through both detectors. That will give you two sets of bounding boxes. You can check which bounding boxes overlap, and by how much using the bboxOverlapRatio function.
What you do next depends on what you are trying to achieve. If you want to reduce the false positives, then you can take only the cases where both methods detect the person, i.e. where there is an overlap between the bounding boxes from each detector. If you want to reduce false negatives, you can take the cases where either method detects a person. Or you can have some kind of weighting scheme depending on which method you trust more.
More Answers (2)
Walter Roberson on 14 Feb 2016
Give a "vote" to each of the detection methods -- not necessarily an equal vote. Any location detected as occupied by both algorithms would get full marks. Any location detected as occupied by only one of the algorithms would get partial marks. Make a decision based upon the scores at the location.
For example, suppose Viola is really good at finding a human face but not good at finding the outlines of the body, and suppose HOG is good for finding outlines but is not good at determining whether there is a face. Then when you applied both algorithms together you would find areas that the detection overlapped. You could then use those as "seeds" to grow the regions out to the largest surrounding area detected by either algorithm.