I'm doing preliminary investigation into an automotive (vision based) pedestrian warning system. from what I can see, this is typically done with motion based tracking in IP cameras. the problem therein obviously being that this will not work on a moving platform. So for vehicle mounted pedestrian detection one would have to use a Histogram of orientated gradients (HOG) based computer vision algorithm, or something similar.
My problem is that:
1. HOG seems to be seldom implemented in embedded image processing libraries(which seems to be the case with TI's wonderful libs too)
2. HOG is EXTREMELY processor intensive. in an initial test using openCV on my quad core I7 PC, it took ~600ms to perform HOG based detection on a VGA sized image!
I have seen some efforts to use openCV on embedded linux based systems to some degree of success, even murmurs of TI porting it to the Davinci platform. but even so I'm not convinced they would be fast enough. so i'd like to know if anyone has ever tried implementing such an algorthm (and if they did, where I could find it). even if you haven't, i'd still appreciate an informed opinion of the matter.
The only other embedded options I could think of would be:
1. something like Toshiba's Visconti platform, which is purposely designed for such systems and has a fast processor with hardware HOG blocks.
2. A FPGA SoC like the Zync platform where one might be able to accelerate the HOG algorithm in hardware blocks. (this would be the harder, more expensive option of course)