Take camera footage from pedestrian overpass (North Avenue) and count the number of vehicles that pass.

Methodology:

Initially I had expected that using optical flow and implement the following logic:

1. identify motion "blobs" between frames
2. a motion "blob" that exceeded a certain threshold size would be a car
3. compute centroid of motion "blobs"
4. when centroid passed through plane of image, the number of cars seen would be incremented by one

What actually transpired:

The following link shows a 30 second clip of traffic passing underneath the North Avenue pedestrian walkway:  30 second clip of traffic flow (compressed)

Optical flow, rather than generating simple motion "blobs", takes the above 30 second clip and turns it into a motion "cloud": optical flow generated from 30 second clip of traffic flow

After playing around with morphological operators and BW thresholding, I still wasn't able to clean up the motion "cloud" to a point that would provide consistent results:  30 seconds of optical flow after applying BW threshold & morphological operators

The best explanation I have come up with to explain away my optical flow problems, is that the original footage was taken from a camera that was not kept 100% stationary.  The tripod I had planned on using to stabilize the camera initially was not tall enough to "see" over the cement handrail of the pedestrian overpass.  As a result, the human bipod was used to stabilize the camera, with less than perfect results.

I strongly suspect that in order for optical flow techniques to be useful, I need to encode a basic "digital stabilizer" to steady the image and prevent it from moving around excessively.

Methodology rethought:

Since the motion "cloud" data obtained from optical flow techniques was so difficult to handle, I stepped back and used the simple technique of image differencing to identify motion.  This time, things went much more smoothly:  30 second clip of traffic flow processed with image differencing

This technique provides the motion "blobs" that I had originally intended to derive from optical flow data.  After simple thresholding, the strength of image differencing for this application was obvious:  BW thresholding of 30 sec image differencing data

Things to come:

Overall progress has been occurring at a phenomenonly slow rate.  I am hoping that this is due to the learning curve associated with Visual C++ & the need to develop tools (i.e. morphological operators, matrix manipulation tools, etc) as problems present themselves.

For now, I plan on continuing to approach the problem using image differencing data.  However, I am rather curious about the feasibility of "digitally stabilizing" the poor camera image which should allow the use of optical flow instead of image differencing to obtain motion information.  Yet I suspect that the general lack of texture variations when looking at a road may significantly hinder any such attempt.