I worked with mmWave radars previously, but not professionally. In my opinion, it would be quite challenging to extract enough features from the current generation of mmWave radars, such as TI IWR1642 ([1]) which have around 4 receivers and 2-3 transmitters, because the incoming data, while having a lot of temporal resolution is very limited in spatial resolution.
With greater number of antennas something like what you describe is theoretically possible, but becomes cost prohibitive.
By limiting the number of transmitters / receivers, we can have an almost perfect privacy monitoring. For example, it would be possible to detect a attack / a fight in a public bathroom, while not exposing anyone.
To an extent, but depending on the sampling rate and frequency and your ability to control the area being observed, there is still a lot of information available for modeling and biometric identification. In the extreme you can detect things like heartbeat, rate of breathing, etc.
For example, I have a CDM324 24Ghz radar module here on my desk. I set it up to 'watch' me type this comment from across the keyboard. This is an extremely simple module that I have powered by a bench power supply and the IF routed directly to the audio input on my desktop. It was sampled with audacity and amplified a bit to help with visibility. Towards the end of the recording I'm expecting a flat spot followed by regular motion followed by 'noise' as I pause motionless for ~10 seconds or so, then take about 5-6 exaggerated breaths, then resume typing.
This is with zero design, the wrong frequency for the job, and next to zero signal conditioning.
(post: i've included a zoomed in image of the 'motion demonstration' to show still/breathing/typing, then zoomed in on typing to show the detail, then a spectrogram and waveform of me reaching up to scratch my ear.)
The difference is the number of transmitters / receivers. TSA uses an array of them (from top to bottom) with narrow beams and they also rotate them and ask you to not move, so that effectively their number is considerably higher (ten of thousands), when reconstructing the image. Additionally, they use a higher frequency (160-400GHz), which additionally helps with resolution. See [1] and [2].
If we limit the number of transmitters / receivers to 3/4, the image reconstruction becomes impossible, similar how you would not really make a real photo from a few color sensors stationary looking at a scene, while a linear array of them that moves around an object would make a perfect scanner. It's a non-ideal analogy, but the best I can offer.
This is true in the sense that NNs offer a convenient way to estimate a distribution of hidden values conditionalized upon visible values. That is an approximation of what is usually understood by "a reconstruction". I would argue that a reconstruction per se would only include parts for which the credible interval could be systematically bounded within a suitable tolerance threshold.
Google knowing every time I’m in the bathroom and keeping a detailed log of the times is a far cry from someone in my family happening to notice I go in there once.
I have no opinion on your reply, I was discussing technical feasibility in response to someone claiming it was infeasible. The normative claims aren't germane or particularly interesting to me.
By not exposing anyone I meant reconstructing an image. Identification is a different story and there are multiple ways on how to achieve that (heartbeat classification is an obvious one).
If you have ideas how to reconstruct an image with 4 receivers / 2-3 transmitters, please, buy a dev kit from TI (they cost <$200, https://www.mouser.com/c/?q=mmwave) and show us a demo.
I have a lot of experience with TI's IWR6843AOP chip and dev kit. It has 3 transmitters and 4 receivers and it'd be impossible to create a facial reconstruction with it.
I've just used it for the point cloud data which can be streamed via USB. If you want to stream the raw data I think you'll need https://www.ti.com/tool/DCA1000EVM. it has a 1Gbps ethernet port on it.
What about correlation? If you have other information that can tell you who is alone in the bathroom (common area CCTV, log of RFID key usage, etc) and your radar can infer that something very private is taking place there because the person thinks they are alone, then isn't that still a huge breach of privacy?
I really must have phrased my message more clearly. I meant reconstructing images, not identifying who is inside. The latter is quite feasible, and you're correct that extra data helps with that as well.
Just like with browser pinning, we have a variety of signals, which combined together give a very high rate of correct identification.
With greater number of antennas something like what you describe is theoretically possible, but becomes cost prohibitive.
By limiting the number of transmitters / receivers, we can have an almost perfect privacy monitoring. For example, it would be possible to detect a attack / a fight in a public bathroom, while not exposing anyone.
1. https://www.mouser.ec/datasheet/2/405/1/swru521c-1954464.pdf