Privacy Preserving AI filter

20-April-2026 TLDR: an AI Filter that detects movement and a person's body posture while still preserving thier privacy. The requirements for this came out of a project for Mass Personalization of healthcare , a pre-initiation project. I was playing anyway with things like yolo , pose detection and depth estimation AI models and wanted to see if it is possible to create an AI filter that can preserve privacy but also still be useful, like still be able to tell, the person shape, pose and action they are doing. Requirements

• KISS : keep it simple stupid !! The simpler it is the simpler would be to implement , deploy and run in the wild
• No Complicated devices I wanted to be able to run on any device, and no need of external sensors , like heat sensors, or noise, ect...
• Python : to be able to run in python to be able to be deployable by anyone with some common knoledge of python + js for front side
• Mass personalization healthcare : this ment would love to tackle an illness that is common to many people.

As I wasnted to keep it simple, and no aditional devices, then that meant would be able to run on a spare mobile phone that people would have it. Moreover, from mass personalization point of view it would make sense that an illness like Dementia or Alzheimer's disease would make sense.

For the proof of concept I've chosen to detect a scene , like a person with dementia that goes to gas stove , puts a pot full of water and forgets about it.
For this scene, you need to detect the gast stove, the pot, and the person, while for the person employing a privacy filter (privacy protecting ai filter).
Once the person is detected leaving the scene after the stove is on, after a certain amout of time an alarm or notification can be sent to a designated number.

I've chosen to use camera on mobile phone to detect the pose of the person and also use depth anything 2.0 to detect the depth of the scene and another small yolo model to detect the pot and the stove. As this is using a mobile phone, that means this can be used by anyone, anywhere, with a spare mobile phone. As this is just a proof of concept, (TRL 3-4) no need to go further at this point in time. Results Not to bore you with more details, this is the result (click, move the slider): As you can see from the video on the left is the real video, on the right is the result. from the left you can :

• Detect the gas stove (yolo11m)
• Detect the Pot (yolo11m)
• Detect pose of the person (yolo11m-pose)
• Shape of person / annonymization / scene (depth anything 2.0)

As you can see from the right video after the person is leaving the scene, and the gas stove is on with the pot, tehn a warning is triggered and then an alarm afer a set time.

There are 3 models here running almost in real time: 100ms-200ms. First is the yolo11m model for detection of pot and gas stove (this is a little bit overkill - could have use an 11 nano model and it would have been much faster as accuracy is not an issue for this scene) Second an yolo11m-pose, same as accuracy is not an issue could have used a smaller model like 11nano Thirdly used a Depth Anything model 2.0 ( not the 3.0) to estimate the depth of the scene and person shape. This third model is not really needed, but it is mainly prettyfy. Only the first model needed training on what a gas stove is and what a pot is.
I will build a tutorial in the future on how to do this by yourself if there is interest in this (if this page is visited more then 50 times).

I also applied the same methodology to other scenes as seen below, click on them to find out. Baby falling out of bed Same methodology, privacy filter with AI, detection of a baby falling out of bed. Person feeling sick (belly pain) Same methodology, privacy filter with AI, detection of person with belly problems. Usually humans touch the areas that hurt us, take a look at the pose detection on right People dancing Same methodology, privacy filter with AI. This could help calculate the amount of physical activity a person is doing inside a nursing home, or similar. Can also be used to detect the interaction between 2 people. Domestic violance Same methodology, privacy filter with AI.