Does Tesla and Comma AI etc classify based on adjacent frames, or does it choose what’s what one frame at a time?
I ask because the article mentions a flashing red light, which is common at rural intersections, and having a light flash is a proven way to grab a humans attention because we’re sensitive to the slightest change in brightness: think of a lost hiker flashing a mirror at a rescue plane overhead.
Without being sensitive to the blinking lights designed for humans, the car sees a floating red orb in one frame, and nothing in the next.
This quote from the Reuters article from yesterday makes me believe something about the tech stack makes it difficult to detect patterns of flashing light:
> In February 2020, Tesla's director of autonomous driving technology, Andrej Karpathy, identified a challenge for its Autopilot system: how to recognize when a parked police car's emergency flashing lights are turned on. "This is an example of a new task we would like to know about," Karpathy said at a conference.
> “I was driving and dropped my phone,” Mr. McGee told an officer who responded to the accident, according to a recording from a police body camera. “I looked down, and I ran the stop sign and hit the guy’s car.”
What a fucking idiot. Handsfree exists for that same use case without holding the phone.
Also where was the "Driver Monitoring System" that was supposed to monitor the driver's attention and their eyes on the road whist operating the vehicle with autopilot on?
I think lots of drivers have become too complacent and relying too heavily on the autopilot system and are still not paying attention when using it.
Maybe if Tesla had put in place a system that properly monitors the eyes of the driver perhaps, the car would slow down by itself and pull over, instead of running past a red traffic light and killing a person.
Tesla only requires periodic, light torque on the wheel while other manufacturers (Subaru, GM I believe?) have moved to eye tracking. Pretty obvious why Tesla doesn’t want to go down that path, they don’t want to invest in a strong implication you have to watch the road.
> Tesla only requires periodic, light torque on the wheel while other manufacturers (Subaru, GM I believe?) have moved to eye tracking.
That clearly didn't work when the driver of that car was distracted and went down to search for his phone when he dropped it. He took his hands off the wheel and the car still didn't take action to detect that he was distracted.
> Pretty obvious why Tesla doesn’t want to go down that path, they don’t want to invest in a strong implication you have to watch the road.
It is also 'pretty obvious' that using the wheel to assess driver attentiveness is not enough, otherwise the car would have slowed down and detected that the driver got distracted.
Tracking the drivers eyes has also worked well for Comma.ai.
Was the occupant of the impacted vehicle wearing a seatbelt? The article does not say. I am just curious. This question implies no stance or statement.
7 comments
[ 0.18 ms ] story [ 25.7 ms ] threadI ask because the article mentions a flashing red light, which is common at rural intersections, and having a light flash is a proven way to grab a humans attention because we’re sensitive to the slightest change in brightness: think of a lost hiker flashing a mirror at a rescue plane overhead.
Without being sensitive to the blinking lights designed for humans, the car sees a floating red orb in one frame, and nothing in the next.
This quote from the Reuters article from yesterday makes me believe something about the tech stack makes it difficult to detect patterns of flashing light:
> In February 2020, Tesla's director of autonomous driving technology, Andrej Karpathy, identified a challenge for its Autopilot system: how to recognize when a parked police car's emergency flashing lights are turned on. "This is an example of a new task we would like to know about," Karpathy said at a conference.
What a fucking idiot. Handsfree exists for that same use case without holding the phone.
Also where was the "Driver Monitoring System" that was supposed to monitor the driver's attention and their eyes on the road whist operating the vehicle with autopilot on?
I think lots of drivers have become too complacent and relying too heavily on the autopilot system and are still not paying attention when using it.
Maybe if Tesla had put in place a system that properly monitors the eyes of the driver perhaps, the car would slow down by itself and pull over, instead of running past a red traffic light and killing a person.
That clearly didn't work when the driver of that car was distracted and went down to search for his phone when he dropped it. He took his hands off the wheel and the car still didn't take action to detect that he was distracted.
> Pretty obvious why Tesla doesn’t want to go down that path, they don’t want to invest in a strong implication you have to watch the road.
It is also 'pretty obvious' that using the wheel to assess driver attentiveness is not enough, otherwise the car would have slowed down and detected that the driver got distracted.
Tracking the drivers eyes has also worked well for Comma.ai.