As I understand it, the deepfaune's first pass is an object detector was based on megadetector, @schamaille  could explain it exactly. In short though, it's output is standard yolo like in terms of properties. From this I use standard opencv code to snip out the individual matches and pass them to the second stage, which is a classifier.

My code needs a bit of cleaning up before I can release it, also it needs to be made more robust  for some situations. Also, I'm waiting to hear if I got anywhere with wildlab awards as it would affect my plans going forward. And this could be anything up till the end of next month, though at a wild guessing I'm guessing next week at the UN WWD or at the wildlabs get together :) Anyone else have any theories ?

Also, my code is a little  more complex because I abstract the interface to a network based API.

Finally, I don't want to take the wind out of my sails, I would like to launch my integration in time with the release of the Orin based version of my StalkedByTheState software, the usage of which I'm trying to promote. To release earlier take's some of the oomph out of this.

But maybe we can have a video call sometime and we can have a chat about this?

Absolutely! I pretty much do the same thing, the resizing step I think relates to what I still have to do. Some large images caused my code to crash.

I want to take it one step further, and that's one of the reasons I want to talk to Microsoft about, I'd like to encourage the abstraction of the object detection with the network API approach I developed as that would mean that any new models anyone developed would simply work out of the box with no additional work with my video alerting software. To that end I need to have a chat to see if they agree with the added value, if so they could potentially add this wrapper around their code and all of those models would be available to alert on and to use is simple Python scripts in other peoples pipelines.

Anyway. That's the plan.

Hi Kevin and Kim, 

Apologies for the delayed response; however, we were unable to provide specific information until today. We aim to reach out to each applicant in the upcoming weeks or months.

All the best, 

Great suggestion! Diving bird studies and their analyses are actually what has helped me get thus far with solving this problem. They happen to have done quite the same thing as I'm trying to do, just with more behaviors added. I believe the study was done with murres and kittiwakes.

Best,

Travis

I'm very close to solving the problem. Just waiting for a function to run on a fairly large dataset to see the results. I will share the repository link with you when it gets accomplished!

The species I'm working with roost atop cave ceiling as also drop from there to get airborne!

Yes, they are triaxial (Technosmart) and body mounted right on their backs.

So far, I have created thresholds for different metrics derived from the accelerometer data. Essentially, I sectioned out a bunch of ACC data where I am positive flight is occurring, and did the exact same with roosting, and crawling around/scratching(activity while roosting). From there, I plotted the distribution of all the metrics to see which metrics have unique distributions that were significantly different than roosting/activity.

Using those distributions, I created thresholds for the important metrics in which all flight behavior was either above or below a certain value for that metric. This got me to being able to construct a decision tree based on these metrics which had pretty solid accuracy.

The downside is a small chunk of flight from the beginning and end of flight bouts are not being included in the behavior classification. I noticed that their wbf during those small chunk are indicative of flight and am going to try and add wbf as the last decision on the tree to improve the accuracy of it.

VeDBA is also being included and calculated and based on the values for the thresholds I have created for flight it should not matter how high their head is, rather how low it is, when x y and z thresholds are also met. If that makes sense.

Hope I answered most of your questions!

Were you ever able to solve the problem? Interestingly enough, I begin a seal bio-logging study next year!

Also, you are correct. The errors were occurring during short bout flights as well as some spectral leakage, but I may have solved the problem by lower the window size. I've also corrected for the spectral leakage by creating a separate function that identifies any significant changes in calculated WBF that last < 2 seconds, then counts number of heave amplitudes within 1 second. I'm using an fft for the calculations and am just waiting for a function to run on a larger dataset to see if everything came out the way I am hoping for. Fingers crossed.

Best,

Travis

Thanks for helping me!

For now, I'll be testing with mains power (220V). The final prototype needs to be functional using solar panel. 

The main system will be a Raspberry Pi, as the Brain of this project, receiving and storage all the measurements from the other devices, controlling the system. The measurement devices will be a microcontroller (still searching for which one to use) with differents sensor, such Lux sensor like TSL2561 or 2591, AudioMoth to record animals sounds, temperature and some other measurements.

For synchronization, must be effective, because the measurements must be or have the same time, so the storage data from differents devices can be grouped by it's timestamp and have full control for animals and fauna. So, probably it will be a lot of data per day.

The distance between each devices I haven't study yet, consider 20 meters within each device.

Thanks @alexrood for creating this visual!