Autonomous and Semi-Autonomous Vehicles

Autonomous and Semi-Autonomous Vehicles
An Introduction

Prof. David Bernstein
James Madison University

Computer Science Department

bernstdh@jmu.edu

Society of Automotive Engineers (SAE) Levels

Driver Assistance - The vehicle contains some features that assist the driver
Partial Automation - The driver must remain engaged and monitor at all times but the vehicle has some automated features (e.g., adaptive cruise control)
Conditional Automation - The driver must be present and ready to take control with notice (i.e., needn't be monitoring the environment but must be monitoring the vehicle)
High Automation - The vehicle is capable of performing all driving functions under some conditions
Full Automation - The vehicle is capable of performing all driving functions under all conditions

Potential Benefits

Safety:
- Automated systems may make fewer errors than humans and, hence, be involved in fewer crashes and fewer serious crashes
Congestion and Capacity:
- It may be possible to reduce congestion by smoothing traffic flows
- It may be possible to increase capacity by decreasing safe following distances (a.k.a., platooning)
Mobility:
- Many people currently can't drive (e.g., because of vision problems, seizure risks, etc.)
- Autonomous transit vehicles could increase the supply of public transportation

Potential Benefits - Crash Reduction

The Idea:
- Eliminate human errors of various kinds
Limits:
- The Insurance Institute for Highway Safety (IIHS) reviewed about 5000 crashes and divided them into the following errors: sensing and perceiving (24%), predicting, planning/deciding (40%), execution (20%), incapacitation including DWI (10%), unavoidable (2%), unknown (4%)

Potential Benefits - Platooning

The Idea:
- Improved sensing and decreased reaction times may decrease the safe following distance (e.g., from 1 vehicle length for every 10mph to 1 vehicle length for every 30mph)
Limits:
- Maximum healthy deceleration rate
Using the Driver's Rule of Thumb:

History

1980s:
- The Defense Advanced Research Projects Agency (DARPA) funded several projects
1990s:
- The National Automated Highway System (which considered vehicle+infrastructure technologies)
2000s:
- The DARPA Grand Challenges in 2004 (no successes), 2005 (5 successes), and 2007 (in an urban environment)
2010s:
- The private sector got much more interested
- Several states started allowing the testing on public roads

Data

About the Network:
- Street network databases at a very high level of detail and accuracy (including speed limits, turn restrictions, etc.) are very useful
About Other Vehicles/Drivers:
- Databases about vehicle and driver behavior may improve probabilistic predictive models

Sensing the Environment

The Sensors Used:
- Digital Cameras
- LIDAR, RADAR and SONAR
Sensor Fusion:
- Unlike with individual driver assistance technologies (e.g., lane departure, blind spot, collision avoidance systems), an autonomous vehicle system needs to be able to combine the data from all of the different sensors
Communication:
- Connected vehicle networks may play an important role in the future, but currently do not

Algorithms

Optimization and Optimal Control:
- Used for the "simple" problems (e.g., route finding, acceleration/deceleration, steering adjustment)
Custom Algorithms and Heuristics:
- Used for driving-specific problems (e.g., object avoidance, movement prediction)
Machine Learning and Artificial Intelligence:
- Used for object (e.g., pedestrian, vehicle, traffic signal) identification and classification
- Used for the "hard" problems (e.g., which pedestrian to avoid)

Nerd Humor

(Courtesy of xkcd)

Processing

In the "Factory":
- Large networks of computers can be used to train and test data off-line
In the Vehicle:
- Traditional processors
- Vector processors
- Specialized processors

Enhancing the Infrastructure

Lane/Road Markers:
- Magnetic pins embedded in the pavement and/or special paints can facilitate sensing
Local Information:
- Electronic street signs can eliminate/reduce the need for optical character recognition to determine "point" information
- Over-the-air databases containing "area" information (e.g., about regulations, construction, etc.)

Some Unanswered Questions

The Definition of Quality:
- Should we just consider the outcomes (e.g., number of crashes) or the processes?
- Should we just consider the "lagging" outcomes (e.g., number of crashes) or also "leading" outcomes (e.g., not creating hazards, responding to hazards)?
Ensuring Quality:
- If outcomes matter, how should we collect the necessary data?
- If processes matter, how should ML and AI algorithms be trained and tested?
Maintaining Quality:
- Should algorithms be allowed to "learn" (i.e., adapt) after the vehicle leaves the factory
- What should prompt a recall and what will the impacts be?
Liability:
- Who is liable if something goes wrong?

More Nerd Humor

(Courtesy of xkcd)