People can mask USPS AI detections

[rdagger]

I implemented a custom DeepStack model to alert me when the mail is delivered because my mailbox is down the street from my home. I'm using a camera mounted near the street that also detects people and vehicles (including ALPR).
The post truck detection works but will fail if the event is preceded by another event. For example, a jogger triggered an alert at 11:18:29 which then masked the mail delivery 6 seconds later at 11:18:35. Is there a way to ensure sequential alerts fire?

 

[wittaj]

You would have to either make the time a lot less or set up a clone camera that will only alert/trigger for b usps

 

[rdagger]

You would have to either make the time a lot less or set up a clone camera that will only alert/trigger for b usps

Are you referring to the trigger break time? I have it set at 10 seconds. Seems like setting it under 6 seconds could end up truncating events.

Cloning sounds like a decent workaround. Are there any performance hits to creating clones? I'm thinking I should make several clones of the street camera to address triggering issues with the ALPR. Zone crossing triggers that work well for people often fail for capturing a clean license plate image. Thanks for the help!

 

[wittaj]

For my LPR I have it at 0.1s so I get a car riding someone's bumper lol. I still get the whole vehicle.

The clones are great and use very minimal CPU. I have several clones in use. Just make sure that you check clone master for the original camera and that it is unchecked in the clones. Then it isn't pulling the video from the camera twice.

 

[Cameraguy]

I implemented a custom DeepStack model to alert me when the mail is delivered because my mailbox is down the street from my home. I'm using a camera mounted near the street that also detects people and vehicles (including ALPR).
The post truck detection works but will fail if the event is preceded by another event. For example, a jogger triggered an alert at 11:18:29 which then masked the mail delivery 6 seconds later at 11:18:35. Is there a way to ensure sequential alerts fire?
View attachment 133051View attachment 133052View attachment 133053

Can you share this usps model?

 

[rdagger]

Can you share this usps model?

Here's where I got it:

GitHub - sstratoti/DeepStack_USPS: Custom DeepStack Model for detecting USPS logo

Custom DeepStack Model for detecting USPS logo. Contribute to sstratoti/DeepStack_USPS development by creating an account on GitHub.

github.com

 

[Cameraguy]

Here's where I got it:

GitHub - sstratoti/DeepStack_USPS: Custom DeepStack Model for detecting USPS logo

Custom DeepStack Model for detecting USPS logo. Contribute to sstratoti/DeepStack_USPS development by creating an account on GitHub.

github.com

Thanks

 

[VideoDad]

Can you share this usps model?

I created an expanded delivery vehicle model building off of the hard work of sstratoti and with help from mikelud1. The custom model is trained to recognize amazon, dhl, fedex, uhaul, ups, usps.

Update: Fully retrained with the latest set of delivery images and exported to 4 different models (YOLOv5 6.2, YOLOv5 NET, YOLOv8, YOLOv8 NET)

Download link: Custom delivery models from VideoDad's Github

 

[rdagger]

I created an expanded delivery vehicle model building off of the hard work of sstratoti and with help from mikelud1. The custom model is trained to recognize amazon, dhl, fedex, homedepot, ikea, lowes, target, ups, usps, and walmart.

Sample captures: delivery truck images

Download link: delivery.pt (version 1.0 1.1 1.2 1.3)

Update: Uploaded a version 1.1 to the same link. I was getting some false positives with ups so I added more images and did additional training. Newer model is also smaller and faster.
Update: Uploaded version 1.2 1.3 with additional training on images for usps.
Edit: I'm also working on adding uhaul and perhaps garbage (truck), though it seems to be prone to false positives on garbage trucks. If anyone would like to donate some images of garbage trucks, I'll add them and retrain the model.

That's very cool. I'll check it out. So far, my current model is working so I will test it for a few weeks before upgrading to your version. Thanks.

 

[Futaba]

I created an expanded delivery vehicle model building off of the hard work of sstratoti and with help from mikelud1. The custom model is trained to recognize amazon, dhl, fedex, homedepot, ikea, lowes, target, ups, usps, and walmart.

Sample captures: delivery truck images

Download link: delivery.pt (version 1.0 1.1 1.2 1.3)

Update: Uploaded a version 1.1 to the same link. I was getting some false positives with ups so I added more images and did additional training. Newer model is also smaller and faster.
Update: Uploaded version 1.2 1.3 with additional training on images for usps.
Edit: I'm also working on adding uhaul and perhaps garbage (truck), though it seems to be prone to false positives on garbage trucks. If anyone would like to donate some images of garbage trucks, I'll add them and retrain the model.

Do you need more delivery truck pictures? I can get you some from my cams.

 

[VideoDad]

That's very cool. I'll check it out. So far, my current model is working so I will test it for a few weeks before upgrading to your version. Thanks.
View attachment 133264

Stratotally sent me the set of images he used to train the USPS model. I then added more examples including trucks with open doors, winter and night images, etc. If you are interested in other delivery trucks (e.g. Amazon, UPS, FedEx, etc.), I invite you to try it out. I'm also looking for additional images from the community for any cases that end up giving false positives or false negatives so I can strengthen the model. So if anyone tries it and has suggestions for improvement, let me know.

 

[VideoDad]

Do you need more delivery truck pictures? I can get you some from my cams.

Try the custom model (delivery.pt) as a starting point. If you find the detection isn't strong enough (or is giving false positives for any reason), then yes, I would love more images to add to the model.

 

[Futaba]

Try the custom model (delivery.pt) as a starting point. If you find the detection isn't strong enough (or is giving false positives for any reason), then yes, I would love more images to add to the model.

yep, just cloned a new camera to run delivery.pt. I am waiting for the first delivery trucks to arrive

 

[VideoDad]

yep, just cloned a new camera to run delivery.pt. I am waiting for the first delivery trucks to arrive

In the meantime if you want to play, you can pick a clip with a delivery vehicle and select Tuning & Testing > Analyze with Deepstack to see if the delivery.pt model is active and detecting.

 

[Futaba]

In the meantime if you want to play, you can pick a clip with a delivery vehicle and select Tuning & Testing > Analyze with Deepstack to see if the delivery.pt model is active and detecting.

It looks to be working. It detected UPS and Amazon from this morning and last week.

 

[Futaba]

It has been working pretty well for the most part. It caught UPS, FedEx, Amazon, USPS. I did have a few false positives.

 

[VideoDad]

You might want to add a zone that only covers the area where you might actually see a logo. In particular, exclude the areas with overlay text or outside the road.

I'll see if I can improve the training on "target". It seems to get fooled by circular red blobs because the training images included store signage that was 3-D rather than a flat logo like you might see on a delivery truck. But I'm not seeing actual deliveries by Target anyway so it's not a top priority.

 

[Futaba]

You might want to add a zone that only covers the area where you might actually see a logo. In particular, exclude the areas with overlay text or outside the road.

I am probably not doing this the expected way. this is one of my auto track cameras. It looks over the neighborhood cue-de-sac and follows people and vehicles as they come in and out of the neighborhood. So its view is not fixed.

But I'm not seeing actual deliveries by target anyway so it's not a top priority.

Yeah, i don't care about Target either. I really only care about USPS, UPS, FedEx, DHL, Amazon. Maybe I could get your models and train a different custom model.

 

[VideoDad]

I am probably not doing this the expected way... Its view is not fixed.

I'd still exclude the edges where you might have text from the overlays. Or you might want to clone a camera that sees stuff closer to your house where a delivery truck to your house might be.

Yeah, i don't care about Target either. I really only care about USPS, UPS, FedEx, DHL, Amazon.

You can always just type those in the list of tags to recognize and leave out target and others.

 

[TBurt]

I'd still exclude the edges where you might have text from the overlays. Or you might want to clone a camera that sees stuff closer to your house where a delivery truck to your house might be.


You can always just type those in the list of tags to recognize and leave out target and others.

Thank you ! Seems to work with the few FedEx and USPS recordings from the last couple of days. I also do not really care about Target, but I left in all the delivery objects to see which may or may not give me false alerts. I like how fast and light on resources this model runs. Openlogo is a pig, fat and slow. I only ran it when I knew I was expecting a delivery. This one I cannot really tell it is running, which is great! Thank you again for your time and effort. I will let you know how it does after running for a while.