3
$\begingroup$

We want to set photo cameras in the field and capture images of insects that visit flowers. We want to place these cameras above the flowers and then they should get triggered whenever an insect lands on the flower.

Is also important to find a way to distinguish if is the same insect just flying around and coming back in the next second or so (or a certain time frame). This is important because we would like to have an accurate estimation of how many unique insects visit a particular flower in a certain period of time. Sometimes an insect can create multiple false positives. That is, we can count that insect as multiple individuals, when is actually just the same individual.

I am wondering also how hard is to do this with conventional photo cameras that have incorporated motion sensors.

If this is not the right Stack Exchange site for such a question/discussion, please point me to a more suitable one.

$\endgroup$
2
  • $\begingroup$ You didin't tell which insect you want to capture, but recognition could be quite difficult, due to the fact that they all look like (almost) identical, except their "size"... So at best you can try to identify them from their relative sizes. But their shapes and colors tend to be quite similar. It seems like a difficult task for a visual solution. Nevertheless depending on your camera setup, you can try to build a data base for each insect, and make comparisons based on that. $\endgroup$
    – Fat32
    Jan 12, 2021 at 18:27
  • $\begingroup$ Thanks, @Fat32 for your comment. I do agree is damn challenging. We aim at too many types of pollinator insects, but first, we want to start with higher taxonomic ranks, like orders (Hymenoptera, Diptera, Lepidoptera), and then move down to families for example. $\endgroup$ Jan 15, 2021 at 10:19

1 Answer 1

1
$\begingroup$

My solution would incorporate 3 different components:

  1. Detection.
  2. Recognition.
  3. Tracking.

I'd train an object detector model (Look at YOLO 8 as a starting point) for detecting the insect.
You could use it also for tracking, so in case it flies within the frame you'd be able to safely say it is the same insect.

For recognition you may use many models. The basic starting point would be models which utilizes the Triplet Loss.

There are 2 implicit assumptions in my solution:

  1. The resolution is high enough to have a correct recognition of the insect.
  2. You have enough labeled data set to fine tune the models for this task.

If you have a very high resolution you could also use a wider view on the flower so with tracking you could have high assurance about insect which left are not to be back.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.