10.22-NHTSA_ADAS_Data-061822

Saturday June 11, 2022

SmartDrivingCar.com/10.22-NHTSA_ADAS_Data-061822

22nd edition of the 10th year of SmartDrivingCars eLetter

  NHTSA Releases Initial Data on Safety Performance of Advanced Vehicle Technologies

June 15, Press release, “Today, as part of the U.S. Department of Transportation’s efforts to increase roadway safety and encourage innovation, the National Highway Traffic Safety Administration published the initial round of data it has collected through its Standing General Order issued last year and initial accompanying reports summarizing this data.

The SAE Level 2 advanced driver assistance systems summary report is available here, while the SAE Levels 3-5 automated driving systems summary report is available here. Going forward, NHTSA will release data updates monthly…”   Read more Hmmmm… This is a good start; however, as NHTSA repeats many times, this is just a start and there are many “data limitations”. The most severe may well be the possibility of substantial “sampling bias“, the most severe of which is that each OEM sourced the reported data very differently.  That makes the data between OEMs incomparable.

Also un reported is any measure that would enable a “crash rate” for an OEM to be determined.  One only has a numerator value but no denominator value.

Finally, 392 crashes of “Level 2” cars were reported during the “10” month period of July 2021 and May 15, 2022.  About 12 million vehicles are involved in traffic crashes every year among the 283 million vehicles that operate in the US.  Assuming any one vehicle is unlikely to be involved in more than one crash per yer, it means that each vehicle, on average is involved in 12M/283M =  0.0424 crashes per year.  Thus, if these ADAS cars were involved in crashes at the average rate, and had their ADAS on all the time, the 500 vehicle crashes per year contained in these data would expect to be generated from a fleet of only about 11,800 vehicles (or 0.0042% of the vehicles (“everything being equal”, ADAS on all the time.). 

Consequently, either, …

• These system outrageously reduce crash probabilities, and/or

• maybe some, but we’re probably not much luckier.

• very few of the cars in use during that “10” month period had Level 2 capabilities, and/or

• unfortunately, the VIN number doesn’t identify these cars and only Tesla announces how many sold (I may have missed the reportings)

• very few of the drivers of those cars rarely engaged the Level 2 features, and/or

• likely.  Only Tesla releases data on the utilization of its level 2 features but does so only in aggregate terms that don’t allow for correction of sampling bias associated with engagement in “easy” driving conditions versus “challenging” driving conditions.

• enormous under counting

• likely, only Tesla has the opportunity to either “know all” or sample effectively because of their OtA monitoring of its vehicles.  Everyone else has conveniently kept their heads in the sand. Mercedes didn’t report any; however, during that period I think my Intelligent Cruise Control and Lane Centering were engaged when I hit a deer.  Mercedes must not have been watching me, I didn’t report it and I didn’t get the memo that informed me to do anything. 

Anyway.  It is a start and at least to me the numbers are not startling.

What needs improvement is sourcing of the incidents.  Maybe OtA should be mandated.  At minimum, the VIN should specify the existence of theses capabilities.  Then normal police reportings can begin to “automatically” access the “black box event recorders” (see also Accident data recorder and NHTSA) that are in most cars today.  Unfortunately, privacy concerns makes this not-easy.  So here we are.  It wont be easy to do much better, but we should continue to try. 

What the data do point out is that a substantial number of the crashes involved the rear ending of a stationary object.  I have pointed out repeatedly that the source code of these systems explicitly disregard stationary objects in the lane ahead. Justifying this explicit process is that current sensors  incur unacceptable false positives when trying to determine if sufficient headroom exists under detected stationary object in the lane ahead.  Thus, to avoid braking in response to these rare false positives, stationary objects in the lane ahead are all assumed to be “pass under-able”. 

As one drives, one encounters many stationary objects in the lane ahead.  These are readily sensed and precisely located ahead.  Readily sensed are overpasses, signs, tree canopies, traffic lights, …  all of which can usually be readily passed under.  (As can vehicles ahead that come to rest in vehicle-follower mode.  These are not disregarded because one is in vehicle-follower mode.)

But when one is in vehicle-leader mode and one encounters a stationary object ahead, I believe, most, if not all “Level 2” systems disregard that object and assume the car can pass underneath.  So if you are in vehicle leader mode and come over the crest of a hill to be confronted with a stopped object ahead, your system will disregard that object.  Similarly, if the vehicle that you are following changes lanes forcing you to become a leader, any stationary object ahead will be disregarded.    Alain

SmartDrivingCars ZoomCast Episode 272/ PodCast 272 Ed Niedermeyer 

F. Fishkin, June 16, “With NHTSA releasing the data on 392 crashes involving driver assistance systems, we dive into the significance and take-aways with guest Ed #Niedermeyer, author, journalist and co-host of the #Autonocast.   Join Princeton’s Alain Kornhauser & co-host Fred Fishkin for episode 272 of Smart Driving Cars.“

Technical support provided by: https://www.cartsmobility.com/