AI People counting is a work in progress

[Airwolf]

I think the part DJI is saying isn't too hard is loading someone elses model and using it.

What exactly are you going to try and make a model for if you don't mind me asking?

White-tailed deer.

At the moment it’s a manual process for DJI to load models. You have to send them the processed data, config file used and the pictures used for training. They presumably do some validation, possibly further processing, and then make it available for your drone.

 

[bigbluesky]

I think the part DJI is saying isn't too hard is loading someone elses model and using it.

What exactly are you going to try and make a model for if you don't mind me asking?

Not sounding like it's a sharing platform from what I have heard so far. I was interested in it for counting different livestock. Cows, calves, sheep, lambs, or even different wildlife, like native species such as kangaroos or koala bear

 

[Airwolf]

Not sounding like it's a sharing platform from what I have heard so far.

It seems little bit like a beta right now, so hopefully the future holds more

 

[bigbluesky]

It seems little bit like a beta right now, so hopefully the future holds more

Yeah, hopefully in a few updates things are more refined or made easier. It's probably a good thing I have to wait 6 months to purchase. Hopefully it's past beta stage by then at least and gives a clearer picture of finished product

 

[Airwolf]

Another thought I had .. if I can get this process nailed down I might look at making a web front end other people can use. That way you could just upload images and get the necessary files back, no special knowledge needed.

 

[bigbluesky]

Another thought I had .. if I can get this process nailed down I might look at making a web front end other people can use. That way you could just upload images and get the necessary files back, no special knowledge needed.

Yeah that would be very cool. I was kind of a bit disappointed when I found out the process was a bit beyond me, so something like that would be majorly helpful.

 

[jaja6009]

Here is a Youtube video showing some of the process.
It's long, and has some filler, but shows a lot of the process.

To view this content we will need your consent to set third party cookies.
For more detailed information, see our cookies page.

Accept third party cookies

 

[bigbluesky]

Thanks jaja. Will watch it tonight when I'm on the big screen

 

[bigbluesky]

That was a pretty good watch. Definitely all the mmyolo stuff was obviously a bit much for someone like me. I would definitely be happy to pay to be able to skip that step via software. Hopefully airwolf can come up with something to make it easier for the average person in the future.
Was a great video to understand the current process anyway. I did have to rewind a couple of times as the guy was pretty fast

 

[jaja6009]

Another hurdle is training the model on the 500 to 1000 images. He mentioned maybe using something like Fiver since that would be very labor intensive.

I think we will see another video tutorial on this that is more streamlined from DJI or someone else.

An interesting note is Anaconda on Linux only. I am wondering if you can do this on Windows too.

I would think this is like any other new learning process. The first time, the curve is very steep, and subsequent times go faster and smoother.

 

[Ro3go]

Does anyone know how to increase the recognition capacity of the AI? I get "a message that a maximum of 128 objects can be displayed on the screen.Check data for number of recognition"

 

[Ro3go]

After conducting some tests on persons recognition and having the message of maximum object limits, I realized that by using the AI application found on the SD card and then copying it to the PC and analyzing the photos, other recognition values appear. This detects more people than what the controller visually shows me.

 

[jaja6009]

I attempted to do another crowd counting mission for a concert we had over the weekend.
My Matrice flopped again.
It does not count people in dense crowds well at all.

I found that Google Gemini with the same images does a much better job and was actually somewhat close to the count the promoters gave.

Matrice 207 People!! Not even a good try at this distance from target.





Google Gemini estimated 1500 to 2000. I do not know what the true count was at this time though.

The Matrice is better with people spread out, but still missed large numbers of people in the more crowded areas.




On the above image Gemini gave an estimate of 1200 to 1800.

 

[Ro3go]

I took this image at 60 meters Alt , 217 meters distance, with zoom 8.9x then opened with DJI Ai tools

 

[bigbluesky]

I attempted to do another crowd counting mission for a concert we had over the weekend.
My Matrice flopped again.
It does not count people in dense crowds well at all.

I found that Google Gemini with the same images does a much better job and was actually somewhat close to the count the promoters gave.

Matrice 207 People!! Not even a good try at this distance from target.

View attachment 31136



Google Gemini estimated 1500 to 2000. I do not know what the true count was at this time though.

The Matrice is better with people spread out, but still missed large numbers of people in the more crowded areas.

View attachment 31137


On the above image Gemini gave an estimate of 1200 to 1800.

Jeez. I thought it would be more accurate than that. I mean if its not accurate its not worth much for any thing more than estimates. That's disappointing.