Transcript slides - Department of Computer Science

Driver Behavior Models
NSF DriveSense Workshop
Norfolk, VA Oct 30-31
Mario Gerla
UCLA, Computer Science Dept
The Challenge
Safety
• 33,963 deaths/year (2003)
• 5,800,000 crashes/year
• Leading cause of death for ages 4
to 34
Mobility
• 4.2 billion hours of travel
delay
• $78 billion cost of urban
congestion
Environment
• 2.9 billion gallons of
wasted fuel
• 22% CO2 from vehicles
Will Driver behavior help?
• Can driver reaction models help reduce
accidents?
• Can expected driver compliance help plan
optimal routes, green waves and alternate
transport modes?
• Can the knowledge of driver habits help
plan pollution reduction strategies?
Autonomous Vehicle Control
How much human control? Can drivers go to sleep?
V2V for Platooning
Are drivers prepared to take over in case of attacks?
V2V and cruise control to avoid
Shockwave formations (INFOCOM 14)
VDR = Velocity Dependent Randomization: normal drive
PVS = Partial Velocity Synchronization: advanced cruise control
Intelligent navigation
• GPS Based Navigators
• Dash Express (came to market in 2008):
•
Synergy between Navigator Server and City Transport Authority
NAVOPT – Navigator Assisted Route
Optimization
• On Board Navigator
F5
– Interacts with the Server
– Periodically transmits GPS and route
F2,5
– Receives route instructions
• Manhattan grid (10x10)
– 5 routes (F1~ F5) from source to
destination
– Link capacity: 14,925 [vehicles/h]
• But, will drivers comply?
F2
S
Shortest path
F1
F3,4
F3
F4
…
…
…D
…
…
Analytic Results
Total average delay (h/veh)
0.45
0.4
0.35
Average delay (hour)
0.3
0.25
shortest path
flow deviation
0.2
0.15
0.1
0.05
0
13500 13600 13700 13800 13900 14000 14100 14200 14300 14400 14500 14600 14700 14800
V2V for Safe navigation
• Forward Collision Warning,
• Intersection Collision
Warning…….
• Platooning (eg, trucks)
• Advisories to other vehicles
about road perils
– “Ice on bridge”, “Congestion ahead”,….
V2V communications for Safe Driving
Vehicle type: Cadillac XLR
Curb weight: 3,547 lbs
Speed: 75 mph
Acceleration: + 20m/sec^2
Coefficient of friction: .65
Driver Attention: Yes
Etc.
Vehicle type: Cadillac XLR
Curb weight: 3,547 lbs
Speed: 65 mph
Acceleration: - 5m/sec^2
Coefficient of friction: .65
Driver Attention: Yes
Etc.
Alert Status: None
Alert Status: None
Alert Status: Inattentive Driver on Right
Alert Status: Slowing vehicle ahead
Alert Status: Passing vehicle on left
Vehicle type: Cadillac XLR
Curb weight: 3,547 lbs
Speed: 75 mph
Acceleration: + 10m/sec^2
Coefficient of friction: .65
Driver Attention: Yes
Etc.
Alert Status: Passing Vehicle on left
Vehicle type: Cadillac XLR
Curb weight: 3,547 lbs
Speed: 45 mph
Acceleration: - 20m/sec^2
Coefficient of friction: .65
Driver Attention: No
Etc.
Existing sensors are about
External Probing
• Radio Channels
–
–
–
–
DSRC
WiFI (V2V and V2I)
LTE; LTE Direct
White Spaces
But, radio channels can be attacked!
Autonomous vehicles currently use:
• On Board Sensor Channels
–
–
–
–
Laser, Lidar
Video Cameras
Optical sensors (reading encoded tail light signals)
GPS, accelerometer, acoustic, etc
What about probing driver in the car?
• Driver Behavior important for efficient and safe
navigation:
• A- Compliance models
–
–
–
–
Will driver comply with navigator instructions?
Will driver wait for Green Wave?
Will driver accept congestion fees?
Speed limits?`
• B- Reaction Time models
– Can driver react fast enough to shockwave alerts?
– Reaction to platoon accidents?
• C- Autonomous Car Driver models
– Can the car estimate how long it will take to regain the attention of the
distracted driver?
• D. Physical Conditions Models
– Detect sleepiness, predict medical situation etc
How to build driver behavior model?
• Vehicle monitors the driver:
– Collects from CAN bus relevant signals (brakes, accelerate, steer,
etc)
– Body movements (video camera, kinect, etc)
– Internal activities (music, phone calls, smoking, etc)
• Vehicle monitors other drivers and road traffic:
– Correlation of driving behavior with external traffic
• Vehicle builds a model of the driver
– Use machine learning techniques
How is the driver model used?
• Autonomous vehicle uses the model to determine best
action to avoid accidents:
– Wake up driver or act directly on breaks?
– Mimic driver behavior in autonomous driving
• Traffic authorities use aggregate models for planning
– Aggregate model (for given age group, profession, place of residence, etc)
used to evaluate:
• Congestion fee policies (for example)
• Multimodal transport solutions
• Road access control
– Privacy issue preserved by large number aggregation