Calibrating our computer vision

Where real world coordinates meet a pixel on an image.

At Real Time Traffic, we use computer vision extensively to capture data used by traffic engineers and policy makers to make our roads safer. Computer vision allows us to extract many traffic safety variables from one source.  For instance, from one video feed, we can analyse how different road users (pedestrians, cyclists, motorcyclists and drivers) are using the same stretch of road.


Extracting accurate data that can inform decision making – often in places where internet connectivity is scant or non-existent – comes with its own set of challenges. After analysing more than 20,000 hours of video, we have learnt what works to give us results.


Apart from all the challenges of getting our self-powered edge computers (belovedly named RealiteHubs) to operate continuously for weeks under such harsh conditions, we have developed a few processes to calibrate real world attributes to what we extract from our video. Our target is almost always a stretch of road with live traffic. With that comes a set of safety risks for our field operators. For our computer vision to produce good, reliable results, we need to feed our algorithms with parameters about the camera settings and the real world.


Camera lens distortion

Camera lens is one of the key factors affecting measurement accuracy.  A wide angle lens will capture more of the scene but exaggerate the relative size of objects relative to their distance to the video camera.





Number of pixels on an image to 1 metre in the physical world

Markers in the physical world are used to calibrate pixels on images to actual physical distances. We measure physical distances on known objects on the road side (such as power poles or traffic signals). Sometimes, markers are spray painted on the ground at measured distances and then captured in the video camera as well as the drone image.





A pixel on an image and corresponding GPS coordinates in the physical world

We use a drone to capture GPS coordinates at each of the marked points.  The drone captures images at an angle and viewpoint that allows us to construct a 3D model of the area.





Comparing speed from computer vision to actual speed

And finally, to validate the speed extracted from our computer vision, we drive a floating vehicle past the camera at a known speed.

No Comment

No Comments

Post a Comment