A Quick and Dirty Way to Count Bees

01 May 2020

Today we'll explore some basic image processing techniques to count bees. These require no labeled training data, so we'll try them first before moving on to fancier techniques.

Generally speaking, the number of bees at the entrance of the hive is a good indicator of the health of the hive. This means that the bees are collecting food (nectar and pollen). So let’s try to count some bees with as little effort as possible!

Say I have a video of the entrance of a beehive, like this:

If I wanted to count the number of bees in each frame, I could set up and train an object detection CNN.

But here’s a lazy way to do it.

First, collapse the color data and apply a Gaussian blur to each frame, like so:

Then apply pixel thresholding to get the dark spots, like so:

To interpret the list of points as separate objects, we can use a clustering algorithm. Here’s an example using hierarchical clustering.

It kind of works! From here we can just count the clusters to get an approximation of the number of bees in frame.

Some clear problems though:

I had to determine a few values experimentally, like intensity threshold and clustering distance threshold
This only works because the bees’ heads and thoraxes are the darkest things in frame. Any dark black objects on the ground would also be counted as bees.

This technique is fast but not reliably accurate. But the small amount of computation required makes it suitable for a low-power application.