Our face detection technology

goCount uses camera mounted on a screen, to detect faces and estimate their dwell time.

This approach requires really good face detector, that provides:

  • stable detection in unrestricted environment (various lighting conditions, face pose and etc.)
  • stable detection at a distance of 10 meters from the camera, to count all viewers
  • low CPU usage, to host both goCount and content management software (e.g. Scala) on the same machine
  • high accuracy

We’ve tested several open source and commercial solutions and no one met all the requirements.

That is why we’ve developed our own face detection, which is:

  • stable to lighting conditions and face pose
  • highly accurate (false alarm rate is about 3%)
  • capturing faces at a distance of 0-10 meters from camera
  • requires less CPU
goCount vs. OpenCV

Let’s see how goCount and OpenCV handle the same video.

Resolution: 640×480
Face size range: 24x24px – 480x480px, which is equivalent to 0-5 meters distance
Machine: Intel Core i7-4500U 1.8GHz

OpenCV(lbpfrontalface) goCount
True detections 3/3 3/3
False detections 27 0
Frame processing time 60 ms 30 ms
face tracking not available available
age and gender detection not available available
OpenCV (lbpfrontalface) goCount

As you can see, our face detection is faster and more accurate.