Real time pedestrian detection
i have found a code about real time pedestrian detection and i think it's results seem better than OpenCV's HOGDescriptor.
i have already made simple improvements on the original code and want to work on how to improve it.
if some experienced user test it and share his remarks i will be grateful.
PS: i asked the author about licence of his code, he replied as "We have not specify the licence terms. However, you can use the code for academic and research purposes following the BSD 3-clause one. But you cannot use the code for business purposes."
I'll take a look, but I don't think that's compatible with the OpenCV license. It might be better to write your own implementation of the algorithm. I don't think you can change something to make it yours, at least as far as copyrights are concerned.
@Tetragramm, thanks for your comment. i am supposing what i have done is using the code on research purposes.
As long as you aren't including it in anything you're selling it's fine. But OpenCV is allowed to be included with commercial software, so this can't be part of OpenCV.
Well, it works, and there were very few false alarms. Didn't pick up every pedestrian in my datasets, and was a bit hit and miss on the ones it did find. It seems to depend heavily on the legs. If the legs are off the bottom of the image, or in a strange position, then it misses them. It also slowed below real-time on a 720x576 pixel dataset, and I've got a nice CPU that it was using all the cores of.
Hmm. If I were to re-write this for OpenCV, it doesn't look like there would be anything left of the original code. So it might actually be ok.
Thanks again. your comments are appreciated.
When you say your results are better than opencv can you give more explanations? Did you plot a DET curve? Have you tried on a lot of of test samples? OpenCV log-average miss rate gives me the same results as the INRIA HOG optimised by the authors (46%)
@Romanzo i did not do any comparison about accuracy but it seems there is a significant increase in running speed.
ok that's a good point but for the accuracy you should plot a DET curve otherwise it's hard to be sure which one is the best. Also you can have a look at this paper it's comparing different algorithm. http://www.vision.caltech.edu/publica...
see paper link