The rate of technological advancement on the path to autonomous
vehicles has been rapid, especially over the last 1-2 years. Major improvements
have been achieved at the lower levels of autonomy. Lane departure systems have
progressed from merely issuing warnings to full active Lane Keeping.
Forward collision systems have progressed in performance and scope, offering both active braking and the ability to detect cyclists and pedestrians
as well as other vehicles. However, the range of sensing that a typical human
driver possesses is still far in advance, in many areas, compared to what machines can
achieve, and we are still some way off having sensors that offer the
performance and scope required for SAE Level 4 automation in a number of key
areas. In this week's Insight, we cover some scenarios that are particularly challenging or, as yet, are not
resolved at all.
TECHNICAL CHALLENGE 1: Predicting vehicle braking performance on the road ahead
Forward Collision Warning and Autonomous Emergency Braking
systems need to understand the grip that the road surface will offer when
making decisions to avoid or mitigate collision events. However, road friction
prediction methods are challenging.
SBD’s view of possible solutions
There are two main approaches to estimating the 'grip
performance' immediately ahead of a vehicle.
ONE: On-board approach
TWO: Cloud-based mapping approach
The first approach involves estimating friction by
correlating sensor data (acoustic, temperature, etc) with tire-road
friction parameters or from the dynamic behavior of the vehicle and/or
wheels. The second approach is to build a cloud-based comprehensive friction
map based on inputs from multiple vehicles which continually send friction
information measured by on-board sensors. Once a detailed road condition database
is generated, the information can be fed back to a vehicle in real-time to
support grip and braking models. In practise a solution which incorporates both
approaches and which also utilizes real-time weather data to modify grip will
most likely be required, and considerable machine learning will be needed
before reliability levels suitable for safe autonomous driving are achieved.
Any autonomous vehicle navigating in an environment that
includes pedestrians needs to anticipate the future path of those that are
nearby and accordingly adjust its path to avoid collisions. However, predicting
the motion of human targets is challenging as they tend to obey a number of
loosely defined common-sense rules and social conventions.
Both Stanford and Tokyo Universities have researched the
aspects of pedestrian trajectory prediction using a variety of models,
including Long Short-Term Memory (LSTM) which can learn and predict human
movement. Tokyo University has further developed an encoder-decoder LSTM-based
model which aims to encode motion trajectory and human interaction to predict
long trajectory sequences. Although there is still some way to go, these
predictive models may finally provide a robust solution.
TECHNICAL CHALLENGE 3: Debris on the road – can I hit it?
On seeing debris on the road ahead, most human drivers are
able to make a decision whether to avoid, or drive through, the unexpected item
before reaching it. In this short time-period, the driver is also performing
various risk calculations, weighing the relative risks of stopping, swerving, or
driving through the debris, based on its perceived size, weight and density. An
autonomous vehicle, on encountering debris, will typically follow a four-stage
process before deciding what action to take. The stages involved are:
THREE: Understand item
FOUR: Run through, avoid or stop
Stages ONE and TWO, while challenging, may be achieved using
neural networks that can be trained in real-world situations to detect most
objects that will be encountered. Early identification is preferable, as this
gives the system more time for the subsequent stages. Some additional training
will be needed to assist with classifying less-frequently encountered items (perhaps
a sofa fell off a truck, for example) to complete the classification database.
Stage THREE involves building a complete picture of the object- and will
involve fusion from multiple sensors. For example, if a barrel-type item is
identified, then what is it made of? Is it empty or full? The final stage,
whether to hit or miss the object, is complicated by the dynamic behavior of
the debris. It is very likely to move between initial identification and
arrival, and its movement will depend on a range of factors, including weather
conditions and the behavior of any other vehicles in-between our vehicle and
the nature of the object. The use of modeling techniques similar to those previously
discussed regarding pedestrian trajectories may in the future provide a
solution for this challenge.
TECHNICAL CHALLENGE 4: Managing different weather conditions
Current optical sensing systems do not operate well in heavy
fog, snow or heavy rain. These conditions reduce the range that they can operate at,
or even render them entirely non-operational. In most conditions, a human is
able to drive, as even heavy fog, snow or rain will not normally prevent them from
reaching their destination safely, even if a little late.
Within the last year or so, a potential solution has started trialing. Since RADAR is much less sensitive to extreme weather conditions, it
may offer a solution to this challenge. A ground facing RADAR is being developed to analyze the topology of the ground below the road surface and to compare the results with a database of the road network. It is expected that, in the future, this information may be able to assist
a vehicle to continue driving, and remain in lane on a highway, in all weathers.
“These challenges represent a significant barrier to the safe deployment of SAE
Level 4 vehicles” says Dr. Alain Dunoyer,
Head of Autonomous Research and Consulting at SBD Automotive. “It is frequently
the case that the legal framework is touted as the reason for a lack of
deployment of such advanced autonomy when, actually, many of the key technical challenges are still far
from resolved. Although it may take time and effort, legislation can be
changed but laws of physics cannot, and seeing SAE Level 4 vehicles on the road may take significantly longer than previously assumed.”