Better Background Subtraction


We’re looking for a developer to help refine the computer vision on an existing project. The project involves capturing silhouettes of people from a top-down IR camera, and re-projecting their silhouette as shadow. Video here:

We’ve been having issues with rain and fog triggering the recording mechanism, and would like a general improvement of the imaging, ideally automatically adjusting to new conditions.

Our main priorities are:

  • ignoring atmospheric conditions such as fog, rain and snow.
  • successfully imaging all clothing types and colours
  • automatically calibrating camera based on new

Our current hardware setup uses a PSEye camera, black acrylic as a filter, and an Olson led cluster as IR illumination. We’re open to upgrading the hardware based on recommendations.

If this sounds interesting, please get in touch at

the video looks interesting.
I am busy with my work
if there are enough resources. maybe I can help
first of all it is needed to save video files to simulate the situation.
and analyse the signal that is used for the video record trigger. if the signal is not good to use for trigger. the signal need improvement, maybe combination with other related signals.
it is definite that more resources and analyse will bring better result. not only for the record trigger but also to image processing

hey, thanks for your response. Here’s a video from one of the cameras - it’s low framerate but gives a pretty good idea of the issues:
We’re currently using frame differencing to trigger the recording, which is clearly an issue with rain and fog.
Could you send me an email with an estimate on what your rates would be for this type of project?

For reference, we ended up upgrading our cameras to Point Grey Blackfly cameras, which gave us a bit more contrast and a lot less noise. We also used blob detection at the end of our image processing, which gave us much better fill.