![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
Gait Recognition and AI
Dec. 11th, 2020 11:12 pmA question arose this evening during Fancy Friday about whether anyone is training an AI for gait recognition.
https://arxiv.org/pdf/1811.00338.pdf
Uses smartphone accelerometer and other sensor data (so this is not using video cameras to recognize someone walking). The introduction does refer to some earlier vision based work.
https://www.businessinsider.com/ai-training-beyond-facial-recognition-gait-detection-heartbeat-sensors-2019-10#gait-recognition-identifies-humans-in-video-footage-by-detecting-their-stride-and-the-software-is-already-being-used-by-the-chinese-government-to-monitor-people-1
Refers to Pentagon using smartphone detection, and Chinese police using some vision stuff, and other recognition strategies as well — the butt thing I remember hearing about last year, the laser cardiac recognition is pretty interesting, as is the floor based sensor identification.
I’m only really looking at the neural network stuff, and this nature article attempts to work backwards from what the deep learning neural networks come up with to try to better understand gait. So, that’s interesting, too.
https://www.nature.com/articles/s41598-019-38748-8
“ Hence, the input relevance values point out which gait characteristics were most relevant for the identification of a certain individual.” The article as a whole is pretty opaque, and minimally helpful in that they instrumented the heck out of the subjects, and the discussion we were having was more about trying to identify people in a surveillance environment, rather than a diagnostic environment.
https://www.mdpi.com/2076-3417/10/3/774/htm
Back to sensors in the floor. This one is a little chilling if it actually scales up — they need only a couple of strides, which means you could basically instrument an entry / exit and do all kinds of hair raising things.
This paper is much less opaque than the Nature one. It is about footstep recognition. It is pretty awesome, in that their sensors are not ridiculously expensive, and that while they did not allow high heels, otherwise people wore their normal shoes, making this a shorter step (har de har har) to real world use. Also, they were able to reuse a lot of the math from image recognition, and the cheaper neural network techniques. Adding new people to the trained network was not computationally intense.
There’s a LOT to worry about here.
https://arxiv.org/pdf/1811.00338.pdf
Uses smartphone accelerometer and other sensor data (so this is not using video cameras to recognize someone walking). The introduction does refer to some earlier vision based work.
https://www.businessinsider.com/ai-training-beyond-facial-recognition-gait-detection-heartbeat-sensors-2019-10#gait-recognition-identifies-humans-in-video-footage-by-detecting-their-stride-and-the-software-is-already-being-used-by-the-chinese-government-to-monitor-people-1
Refers to Pentagon using smartphone detection, and Chinese police using some vision stuff, and other recognition strategies as well — the butt thing I remember hearing about last year, the laser cardiac recognition is pretty interesting, as is the floor based sensor identification.
I’m only really looking at the neural network stuff, and this nature article attempts to work backwards from what the deep learning neural networks come up with to try to better understand gait. So, that’s interesting, too.
https://www.nature.com/articles/s41598-019-38748-8
“ Hence, the input relevance values point out which gait characteristics were most relevant for the identification of a certain individual.” The article as a whole is pretty opaque, and minimally helpful in that they instrumented the heck out of the subjects, and the discussion we were having was more about trying to identify people in a surveillance environment, rather than a diagnostic environment.
https://www.mdpi.com/2076-3417/10/3/774/htm
Back to sensors in the floor. This one is a little chilling if it actually scales up — they need only a couple of strides, which means you could basically instrument an entry / exit and do all kinds of hair raising things.
This paper is much less opaque than the Nature one. It is about footstep recognition. It is pretty awesome, in that their sensors are not ridiculously expensive, and that while they did not allow high heels, otherwise people wore their normal shoes, making this a shorter step (har de har har) to real world use. Also, they were able to reuse a lot of the math from image recognition, and the cheaper neural network techniques. Adding new people to the trained network was not computationally intense.
There’s a LOT to worry about here.