Package detection

[brianegge]

I have a little write up on my project to notify me when a package is delivered. So, doesn’t work with AI tool /blue iris yet, but I may train a deep stack model for it.

Package Detection with a Jetson Nano and Computer Vision

How a developer built local package detection with a $100 Jetson Nano and an IP camera: 15 starter images, iterative retraining, no cloud fees.

blog.roboflow.com

 

[Aquaman]

I've been playing with machine learning on a RPI4 streaming RTSP from my cameras that are otherwise doing their usual thing . In my case, I'm detecting people and their dogs who have a desire to leave gifts on my lawn... my sprinklers turn on as they step anywhere on my lawn. For package detection, an RPi3 should be more than fast enough as it's not a moving object. Plus it's low power and doesn't cost anything other than the RPi!!

 

[brianegge]

an RPi3 should be more than fast enough as it's not a moving object

I wouldn't recommend the Raspberry Pi. I have a rpi3 whose job is to look at the garage camera and decided which cars are present and if the garbage bins are there. Using MobileNet it takes it 6-7 seconds to analyze one frame. This is fine as it only does so when a door closing triggers it. However, I wouldn't do it for real time object detection. Nvidia has a cheaper Jetson for $60, which would be quite capable for your task. You may be fine with an off-the-shelf yolo 80 model - especially if you map classes to what they reasonably are. I.e., it might confuse a dog and bear, but you can map them both to dog, because it's unlikely someone will be walking their bear.

BUT you could have a custom trained model, which has two classes: dog and dog_pooping. You could build up your dataset in roboflow ai and then train a custom model for DeepStack.