130 likes | 255 Vues
Benchtesting Driver Support and Collision Avoidance Systems using Naturalistic Driving Data. Shane McLaughlin March 17, 2011. System Development Cycle. New technology or application. Observation of drivers. Crash statistics. Carry-over systems. System Concept. Engineer. Prototype.
E N D
Benchtesting Driver Support and Collision Avoidance Systems using Naturalistic Driving Data Shane McLaughlin March 17, 2011
System DevelopmentCycle New technology or application Observation of drivers Crash statistics Carry-over systems System Concept Engineer Prototype Friends and colleagues On-road Test track Participants Participants Simulator
Timeline as Testing Accrues Driving Time and Mileage Cost Prototype version n Weather conditions Ecological Validity Road types Rare Events Driving styles Vehicle types Time Participants
Benchtesting Strategy • Many vehicle subsystems process inputs from the vehicle, machine vision, driver controls, etc. Wheel speed Turn signal state Display Prototype Pedal position Vehicle Control Range to forward objects Vehicle yaw • Many of these are available directly in naturalistic data sets. • Others can be created in post-processing (e.g., a time-to-intersection “sensor”)
Example from Forward Collision Warning (FCW) Development • We want to look at the performance of different warning algorithms. • We want to evaluate false alert rates. • Three public FCW algorithms modeled in MATLAB. • Naturalistic data collected during real crashes (100Car study) were used as inputs. • Algorithm output was evaluated.
Real data input into Algorithm Models Time Available for Driver Response Kinematic Estimate based on observed values and equations of motion (3) (2) Response Necessary (1) Alert Occurs Timeline of Input Data and Algorithm Vehicle Response Time Movement Time Perception Time Time Crash / Near Crash Reaction Start Risk Present • We want to be able to generalize beyond the involved driver’s response capabilities…
Cumulative Distribution of Reaction Time from Driving Literature (for now). % of Population Reaction Time (s) Estimate of Percentage of Population Able to Respond in Time Available Vehicle Response Time Movement Time Perception Time Time Real data input into Algorithm Models Time Available for Driver Response Kinematic Estimate based on observed values and equations of motion Crash / Near Crash Reaction Start Risk Present
One Event 0.675g
Controlled and Stratified Inputs (Age X Vehicle Type) Sedan SUV Pick-up Age 16-17 Young Sedan Drivers 18-20 21-25 36-50 51-65 Older Sedan Drivers 66-75 76+ Performance measures Prototype Benefits Estimation
Controlled and Stratified Inputs (Age X Weather) Snow Rain Clear Age 16-17 Young Drivers in Snow 18-20 21-25 36-50 51-65 Older Drivers in Snow 66-75 76+ Performance measures Prototype Benefits Estimation
Controlled and Stratified Inputs(Age X Speed) 0-25 mph 25-55 mph 55-70 mph 70+ mph Age 16-17 18-20 21-25 36-50 51-65 Older Drivers in 0-25 mph 66-75 76+ Performance measures Prototype Benefits Estimation
Types of Uses • Anticipating the range of conditions in which a system will be operating • Tuning prototype logic • Evaluating value of additional sensors or different sensor sample rates • Exploring prototype behavior in the “corners” of the design space, including rare events • Estimating frequency of applicable driving conditions • Estimating false alert rates
Summary • Design iterations can be performed quickly before physical prototypes are constructed. • Some testing that would require time, vehicles, drivers, recruiting, weather, etc, can be done “on the bench”. • Benefits estimates can be made with real exposure information.