[AutonomousVehicles] Tesla Self Driving Feature Mistakes Moon For Yellow Traffic Light

[Cornelius Butler]

Hi Fellow Committee Members,
There is a new report about Tesla's Full Self Driving Feature Mistaking the
Moon for a Yellow Traffic Light. Below is the article link and full text.

Article Link:
https://www.autoweek.com/news/green-cars/a37114603/tesla-fsd-mistakes-moon-for-traffic-light/

Article Text:

     Tesla's Full Self-Driving system, offered for purchase or via a
monthly subscription, offers traffic light and stop sign recognition.
    Speed limit and traffic light recognition features debuted as part of
FSD in Spring 2020.
    Tesla has begun offering FSD on a subscription basis, but older Tesla
models may also require a $1000 hardware upgrade.

Tesla's Autopilot driver-assist system has been on the market long enough
to generate plenty of data about its ability to identify (of not) impending
obstacles, from other vehicles to various concrete structures. After
several years on the market, most of the items that tend to confuse
Autopilot on a regular basis have been fairly well studied, if not
completely eliminated from the roster of things that tend to produce
unexpected reactions from the system. Some of the more consequential items
have tended to be road lane markings, since Autopilot has relies on them to
steer itself within a lane. Fire trucks parked in highway lanes while
responding to emergency calls have been another not uncommon foe for the
system.

One issue we hadn't really seen until now—because traffic sign recognition
has not been activated for a significant period of time as a part of
Tesla's Full Self-Driving suite—is the system mistaking the moon for a
yellow traffic light.

A Tesla owner recently posted a video showing the Full Self-Driving system
confusing the Moon for a yellow traffic light, which was prompting the car
to slow down.

Of course, the yellow tint of the Moon could be related to wildfire smoke
in the atmosphere over parts of the US, so perhaps this issue won't be one
of any regularity, but we had wondered in the past about various
semi-autonomous systems' abilities to identify and correctly distinguish
and respond to traffic lights, as such systems have arrived in production
passenger cars. Tesla's system is not based on camera sight alone, as it
also relies on map data of intersection and light locations, with the
system designed to slow the vehicle down for all detected lights.

If anything, we'd be more concerned about such systems correctly responding
to traffic lights that apply specifically to them and cars in their lanes,
as some intersections can be very complex or positioned too close to each
other for traffic light recognition systems to pick out the correct ones.
As we all know, traffic lights can at times be positioned in front of other
more distant traffic lights, or amid a jumble of other traffic signs, as in
the real world intersections can have many different lights applying to
various lanes of traffic. Autonomous developers have also had to contend
with intersections where traffic light positions are either too high and
too close to the front of the vehicle to be seen by the camera-based
systems interpreting them. Various light conditions can also easily
interfere with all camera-based systems, not just Tesla's.

One of the ways in which automakers and autonomous system developers have
sought to bypass this issue entirely is though traffic lights that
communicate with the systems in the cars themselves, via
vehicle-to-infrastructure (V2I) technology. For example, Audi's Traffic
Light Information system, deployed in certain cities since 2016, relies on
real-time signal information from a traffic management system via an 4G LTE
data connection.

"When the light is red, the TLI feature will display the time remaining
until the signal changes to green in the instrument cluster in front of the
driver or in the head-up display (if equipped). This 'time-to-green'
information helps reduce stress by letting the driver know approximately
how much time remains before the light changes," Audi says.

Of course's Audi's system is far from functioning at every intersection, as
it's hardware-dependent.

Tesla's system, on the other hand, relies on visual interpretation of the
lights rather than V2I technology, and doesn't rely on signals to the
traffic lights themselves. It's more versatile, as we have seen, but also
somewhat more prone to misinterpreting them.

Of course, one other concern with various systems continuously
misinterpreting traffic lights, not just Tesla's system, is that if the
vehicle makes braking and acceleration decisions based on erroneous sensor
data, then it can reduce its speed on the highway and possibly prompt a
vehicle behind it to rear-end it. And as much as the Moon can be mistaken
for a yellow traffic light, we have a feeling that the sun could be a much
more frequent culprit of the same phenomenon, in addition to other circular
lights such as overhead street lighting.

-- 
Cornelius Butler
President
Butler New Media, LLC
"Creating A More Accessible World"
http://www.butlernewmedia.com
email: corn at butlernewmedia.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://nfbnet.org/pipermail/autonomousvehicles_nfbnet.org/attachments/20210725/218ff31d/attachment.html>