Doorstep Package Monitoring

In this article, I will go through an application which monitors your incoming packages at your doorstep using AI.


As we are mostly home, due to the current pandemic situation, we shop online for a lot of items from different e-commerce websites and the packages get delivered at our doorsteps. How cool would it be that, if you get notified instantly when packages get delivered to or removed from your doorstep? This is exactly what I will be covering in this article.


My motivation for this project was 2-fold. Firstly, the idea came up when alwaysAI asked if I would like to speak and do a demo at a webinar of theirs on accelerated AI at the edge. Secondly, I received the second place in the OpenCV Spatial AI competition where I worked on a parcel classification and dimensioning problem. I was well versed working with computer vision on cardboard packages. More details about that are here.


I used alwaysAI low code platform for making the computer vision application and Balena for deployment. Balena provides a software platform that helps developers build, deploy, and manage the code that runs on connected devices.

Architecture of the application
How to build an Edge AI Computer vision Application

Data Collection

For this project, I trained a custom object detection model that detected packages. The first step in this process was to create my dataset. I gathered a lot of images of packages how they would look at my doorstep under different conditions like lighting, distance, time of the day, angles, combinations, stacked, etc. The more images you collect, the better your model’s accuracy will be in identifying packages. Some examples are shown below. I trained the model using around 200 annotated images.

Annotated Image with package
Annotated Image with packages

Model Training

Model training is the most time-consuming task after dataset creation. alwaysAI has a model training toolkit in place for doing your model training on your laptop without needing the cloud. If you have a laptop or a desktop which has the Nvidia GPU that supports CUDA, you should be sorted to run the training job locally. I trained for 100 epochs and on my laptop it took around 1.5 hours which was great. There is no need to provision any cloud resources to run the training. I used the mobilenet-ssd architecture which is pretty good for this use case because of its speed and decent accuracy. The model training toolkit uses the TensorFlow under the hood for training the model.


I built this application so that anyone can run this application no matter what hardware they have. Broadly speaking, you can run the application in 6 different ways, based on the hardware you have.

Example setup using NCS2 stick and USB cam

Next Steps

Feel free to fork the repository or send a pull request if you have more ideas with this application. This application sends notification to a mobile app. You could integrate it with another application that sounds a buzzer or turn on a light, or take a picture when packages gets dropped at your doorstep. The possibilities are endless. If you like the project, please let know if you did something with it, I would be very happy to know about it.

Linux Enthusiast, Embedded systems, Quick Learner, IoT Developer

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store