Following a series of Tesla crashes, one subject of fierce debate is whether the cabin camera in Tesla's Model 3 can be adapted for use as a driver monitoring system.
I don’t care about Tesla. To be honest, I am mystified by the legions of journalists, experts and Tesla Fanboys that spend their days on Twitter or LinkedIn arguing about the company and its technologies. Yes, electrification is happening, Autopilot is an interesting take on ADAS and over-the-air (OTA) updates is a really neat idea. However, let’s look at some statistics to put things in perspective.
Tesla reports that it delivered about 250,000 vehicles in 2018, up from just over 100,000 in 2017. Compare this with Lada – the target of various funny, but very vulgar, jokes from my younger days – which sold 360,000 vehicles last year in Russia alone.
To me, caring about Tesla is like caring about car production in Belgium. Which I don’t. No offense, Belgium. Out of global light vehicle production of about 95 million in 2018, Tesla had a market share of 0.26%. In my world, that’s not a market share, that’s a rounding error – and I don’t even get out of bed for rounding errors.
Following a series of Tesla crashes – some of which are suspected of resulting from Autopilot misuse – one subject of fierce debate is whether the cabin camera in Tesla’s Model 3 can be adapted for use as a driver monitoring system (DMS). Since many of the Twitterati evidently also read EE Times, let’s have a look at this in more detail.
(Source: Tesla Model 3 Fan)
The cabin camera is located above the rear-view mirror and, therefore, above the driver’s head. This ensures an unobstructed view of the overall interior and all the occupants, but is poorly positioned to provide detailed information about driver state. Any wide-brimmed hat will completely obscure the view of the driver’s face.
Cameras for DMS use are best positioned either at or below the driver’s eyeline. This line of sight provides optimal availability of core signals such as head pose, eye closure, blink speed, blink duration and eye-gaze. Typical camera sites for DMS include the instrument cluster (BMW X5), steering column (Cadillac CT6), and center-stack console (Subaru Forester).
So, perhaps the camera could be useful for occupant monitoring, but the position is terrible for meaningful driver monitoring. Tesla: 0/1.
Next, let us consider illumination. Camera-based DMS must work reliably at night, in direct sunlight, through sunglasses, and in variable and fast changing lighting conditions (think of driving under a row of street lights). Cabin illumination using visible light is unsuitable, since it interferes with the driver’s low-light vision.
The solution is to use near-infrared (NIR) light – typically 940nm – which requires both NIR camera and LED emitters. In the Model 3 there’s clearly a camera, but there’s no LED emitters. With no NIR illumination, it cannot operate in the dark. Maybe the camera could do something in daylight, but, even then, the risk of false positives is substantial. Tesla: 0/2.
Right about now, Fanboys will be screaming: “He’s so stupid! AI can do E-V-E-R-Y-T-H-I-N-G, right? Didn’t we figure that out, like, at CES, back in 2017?”
AI magical thinking brings us conveniently on to the third issue: Is the video output of the camera even connected to any kind of vision processor? Moving quickly on to the fourth point: What is the frame rate of the camera? DMS needs more than twenty frames per second (fps) to provide any meaningful data about the driver – and state of the art is over fifty. If the camera was specified merely to work as a frame-grabber, then even 1 fps is unlikely.
“OTA! He forgot OTA! It can fix A-N-Y-T-H-I-N-G, right? Just how dumb is this guy?”, wail the Fanboys.
Sure, OTA updates can change the software functionality of a system, but cannot change the physical design characteristics. If the camera isn’t connected to a vision processor, or has a very low frame rate, it is game over.
In summary: On the first two points – which both relate to optical flow expertise – it is a fail. On the last two points – which both relate to system design – the specifics are unknown.
At best, the cabin camera in the Model 3 could be upgraded to be the worst camera-based DMS possible. However, my judgement is that it cannot be used for driver monitoring purposes at all – at least not in any way that will provide a meaningful improvement over the existing steering wheel sensor, which was defeated by an orange.
Anyway, you make up your own mind. I don’t care.
— Colin Barnden is principal analyst at Semicast Research.