Why SEE won't be copied by a new AI startup29 Nov 2020 14:38
Seeing Machines isn't just about Machine Learning. Many thought it was especially a few years back so there were many starry-eyed start ups. That's what lead up to the, 'AV is almost here' myth too.
Start with the camera and lighting - we don't make those, but we have patents in his they are implemented, to ensure they work in low sunlight while driving through trees. We don't own 940mm,but we were one of the first with a good implementation.
Finding the face and eyes in an image - you can learn it through Machine Learning, or you can download open source code to do it for you. How we do it is not patented, but we have been doing it for an awful long time and we have refined the process to be the best. Measuring the pupil position and size and the position of the glints here we are in procedural territory. Lots of maths and geometry and knowledge of eye structures allows you to measure calculate wherr the gaze is directed.
Higher order measurements these are derived over a series of measurements. The ML approach is to show tired images and awake images and let the machine guess the difference. But if you are working from the science and HF, then you are measuring eyelid opening and pupil size and gaze direction, then calculating the speed and the acceleration of them as they change you are looking at patterns of where the gaze is directed and how it reacts to events and how that changes over time.
Now validation of your measurements and how the higher order calculations tie in with years of practical research and development to ensure that you are detecting microsleeps, measuring the drivers state and quantifying deterioration in mental attention.
But, if you want to just measure eyes front, yawn, smoke,ub drink, phone and seat belt then ML is sufficient (assuming you have a large, diverse and well labelled dataset) and you can fight it out with all of the other competitors for the first generation of Chinese devices