Upload
trantu
View
217
Download
4
Embed Size (px)
Citation preview
Themes & trendsKey Focus: Driver & Technology
Key Targets: Safety, Energy, & Mobility
40,000/yearX
$9.6 M/human life
USD 384 billion/year + Injuries
Connected Sensors – Dr.
College of Engineering
Warnings
Sensors in vehicle
BIG DATA
A Simple Picture of CAVs
Automation & impacts Work/residence locations, auto ownership & use (4%)
• Shared MobilityTrip Generation & Distribution
• More trips (aging/underserved), Longer trips• Changing demand generators?
Mode choice • New shared ride modes (Uber, Lyft) (↑)• Drive alone?• Transit (↓)
Vehicle & Infrastructure• Light weighting & Electrification/AFVs (Safety?)• Automated lanes, CMS/HAR, Pavement markings
Route assignment• Efficient routing/platooning• Undetermined cost functions• System optimum vs User equilibrium• Micro-behaviors
Anticipated Changes
Concept Development - Driving Volatility
Driving Volatility:The extent of difference from the norm.
Take-away concept
Sources: • Liu, J., X. Wang & A. Khattak. Generating Real-Time Volatility information, Presented at 2014 Intelligent
Transportation Systems World Congress, Detroit, MI, 2014 • Wang, X., A. Khattak, J. Liu, G. Masghati-Amoli & S. Son, Masghati-Amoli & S. Son. What is the Level of Volatility in
Instantaneous Driving Decisions? Transportation Research Part C: Emerging Technologies , 2015.
Different measures of volatility
Driving volatility
Speed-Based Volatility
Vehicular Jerk-Based Volatility• Longitudinal • Lateral
Acceleration/DecelerationBased Volatility
• Longitudinal• Lateral
Different measures of volatility
Driving volatility
Speed-Based Volatility
Vehicular Jerk-Based Volatility• Longitudinal • Lateral
Acceleration/DecelerationBased Volatility
• Longitudinal• Lateral
Statistical Measures of Volatility• Std.dev• Mean ± X*SD• Mean absolute deviance• Coefficient of variation• …..
• Historical financial volatility
𝑆𝑆𝑆𝑆(𝑙𝑙𝑙𝑙𝑥𝑥𝑖𝑖𝑥𝑥𝑖𝑖−1
∗ 100)
Volatility matrix
Trip-based
volatility
Location-based
volatility
Event-based
volatility
Driver-based
volatility
Volatility matrix
Trip-based
volatility
Location-based
volatility
Event-based
volatility
Driver-based
volatility
A BVI:10%
C DVI:14%
FVI: 4%
E
VI: Volatility index
Volatility matrix
Trip-based
volatility
Location-based
volatility
Event-based
volatility
Driver-based
volatility
A BVI:10%
C DVI:14%
FVI: 4%
E
Location “i”
Volatility matrix
Trip-based
volatility
Location-based
volatility
Event-based
volatility
Driver-based
volatility
A BVI:10%
C DVI:14%
FVI: 4%
E
Location “i”
VI: 34%G
Volatility matrix
Trip-based
volatility
Location-based
volatility
Event-based
volatility
Driver-based
volatility
A BVI:10%
C DVI:14%
FVI: 4%
E
Location “i”
VI: 34%G
Long-term Goal of our NSF Study*• Reduce driving volatility ->> control assists• Increase cooperation between proximate vehicles
to respond to unexpected events• Ideas…
■ Understand microscopic driving volatility using Bayesian and Dynamic Markov Switching Framework.
■ MDP, Inverse Reinforcement Learning, Gossip algorithms
• Understand drivers’ decisions at different speeds* Driving Volatility in a Connected and Cooperative Vehicle Environment: Algorithms for DriverWarnings and Control Assists (Sept 2015 – Sept 2018) NSF Award #1538139
CAV Data – Basic Safety MessagesSAE J2735 BSM standard:• BSMs transmitted using
DSRC-range ~1,000 ft
Source: http://www.its.dot.gov/itspac/october2012/PDF/data_availability.pdf
BSM elements
Part I: included in all BSMs
Time
3D Position (Lat, Long, Elevation)
Speed & Heading
Brake Status
Positional Accuracy
Acceleration
Instantaneous Driving Contexts
Steering Wheel Angle
Throttle position
Part II: optional, not included in all BSMs
Vehicle Size
Event Flags
Path History
Path Prediction
Other optional fields
Red indicates High Priority Data Elements
Delivering improved alerts, warnings, and control assistance using basic safety messages transmitted between connected vehicles (Liu and Khattak)
Accel./Decel. (ft/s2)
Critical Area (Mean +/- 2SD Maximum) Normal Area (Mean Mean +/- 2SD) Extreme-Stable Area (Minimum Mean)
Freq
uenc
y(s
econ
d)Ac
cel./
Dec
el.(
ft/s2
)Speed (mph)
Critical Area (Mean +/- 2SD Maximum) Normal Area (Mean Mean +/- 2SD) Extreme-Stable Area (Minimum Mean) Distribution of Quantified Driving Decisions
Threshold Evolution
Point
LineSurface
Quantification of Driving Volatility
(Equation 1)
(Equation 2)
Data Visualization
Gray dots: Volatile Seconds, outside of the Surface Threshold
Surface Threshold for identifying critical driving moments
Instantaneous Driving Profile onRoad
Warning and control assist based on extreme driving seconds (0.5 seconds above 95th percentile)
Jerk
Acceleration
Speed
Real time Info
Daily driving performance summary (or monthly, yearly)
You drove 55 km todayYour average volatility is 22%Average volatility for other drivers in your area is 12%
Driving Profile
For DriversApplications: Driver Feedback
Khattak A., J. Liu & X. Wang. Supporting Calmer Instantaneous Driving Decisions: Use of VehicleTrajectory Data to Generate Alerts and Warnings, TRB paper # 15-1345. Presented at theTransportation Research Board Annual Meeting, National Academies, Washington, D.C., 2015. IETIntelligent Transport Systems, Under 1st round review.
Analysis of volatility in driving regimes extracted from basic safety messages transmitted between connected vehicles (Khattak & Wali, 2017).
Michigan
Ann Arbor – Geo referenced data
Citation: Khattak, A. J., & Wali, B. (2017). Analysis of volatility in driving regimes extracted from basic safety messages transmitted between connected vehicles. Transportation Research Part C: Emerging Technologies, 84, 48-73.
1
3
2
P(2-2)
P(3-3)
P(1-1)
Prob (1-2)Prob (2-1)
Prob (1-3)
Prob (3-1)
Prob (3-2)
Prob (2-3)
Behavior Conceptualization
1 2
2
1
3
P(2-2)
P(2-1)
P(2-3)
-1 s 0 s
1 s
1+n s
1+n s
1+n s
Associated Factors: Instantaneous driving contexts
Three-Regime Markov Switching Dynamic Regression Framework
Analysis of volatility in driving regimes extracted from basic safety messages transmitted between connected vehicles
For short-term prediction…
Regime 1: 𝑦𝑦𝑡𝑡 = 𝜇𝜇1 + ∅𝑦𝑦𝑡𝑡−1 + 𝜀𝜀𝑡𝑡
Regime 2: 𝑦𝑦𝑡𝑡 = 𝜇𝜇2 + ∅𝑦𝑦𝑡𝑡−1 + 𝜀𝜀𝑡𝑡
If timing of switching is known, then:
𝑦𝑦𝑡𝑡 = 𝑠𝑠𝑡𝑡𝜇𝜇1 + 1 − 𝑠𝑠𝑡𝑡 𝜇𝜇2 + ∅𝑦𝑦𝑡𝑡−1 + 𝜀𝜀𝑡𝑡
However, 𝑠𝑠𝑡𝑡 is not observed i.e. stochastic time series driving process. Formulation of MSDR:
𝑦𝑦𝑡𝑡 = 𝜇𝜇𝑠𝑠𝑡𝑡 + 𝑋𝑋𝑡𝑡 ∝ +𝑍𝑍𝑡𝑡𝛽𝛽𝑠𝑠𝑡𝑡 + 𝜀𝜀𝑡𝑡
Where:𝑦𝑦𝑡𝑡 is the dependent variable𝜇𝜇𝑠𝑠𝑡𝑡 is the regime-dependent intercept term𝑋𝑋𝑡𝑡 is a vector of exogenous variables with regime-independent coefficients 𝑍𝑍𝑡𝑡 is a vector of exogenous variables with regime-dependent coefficientsBoth error terms ~ 𝑁𝑁 0,𝜎𝜎𝜎2 and 𝑁𝑁(0,𝜎𝜎𝜎2)
General Time-series Framework
Markov Chains
Irreducible and aperiodic Markov chain from ergodic distribution:
𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃(𝑠𝑠𝑡𝑡+1=1|𝑠𝑠𝑡𝑡=2) = 𝑝𝑝21
𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃(𝑠𝑠𝑡𝑡+1=2|𝑠𝑠𝑡𝑡=1) = 𝑝𝑝12
𝑠𝑠𝑡𝑡 is not observed. 𝑠𝑠𝑡𝑡+1 depends on current state st and not on previous history st−1, st−2, … . .
𝑷𝑷 =𝑝𝑝11 𝑝𝑝12𝑝𝑝21 𝑝𝑝22
Likelihood function with latent regimes
Pr 𝑠𝑠𝑡𝑡 = 𝑖𝑖;𝑦𝑦𝑡𝑡;𝜽𝜽 =)𝑓𝑓(𝑦𝑦𝑡𝑡|𝑠𝑠𝑡𝑡 = 𝑖𝑖, 𝑦𝑦𝑡𝑡−1;𝜽𝜽)𝑃𝑃𝑃𝑃(𝑠𝑠𝑡𝑡 = 𝑖𝑖;𝑦𝑦𝑡𝑡−1;𝜽𝜽
)𝑓𝑓(𝑦𝑦𝑡𝑡| 𝑦𝑦𝑡𝑡−1;𝜽𝜽
In simplest case i.e. constant-only model, 𝜽𝜽 = [𝜇𝜇1,𝜇𝜇2, 𝜎𝜎12, 𝜎𝜎22,𝑝𝑝1−2,𝑝𝑝2−1 ]
L(𝜃𝜃) = ∑𝑡𝑡=1𝑇𝑇 𝑙𝑙𝑃𝑃𝑙𝑙𝑓𝑓(𝑦𝑦𝑡𝑡|𝑦𝑦𝑡𝑡−1;𝜃𝜃)
Transitional Probabilities
𝑝𝑝𝑖𝑖𝑖𝑖=exp(−𝑞𝑞𝑖𝑖𝑖𝑖)
1 + exp −𝑞𝑞𝑖𝑖1 + ⋯+ exp(−𝑞𝑞𝑖𝑖𝑖𝑖)
𝑝𝑝𝑖𝑖𝑖𝑖 = 11+exp −𝑞𝑞𝑖𝑖1 +⋯+exp(−𝑞𝑞𝑖𝑖𝑖𝑖)
𝑗𝑗 ∈ 1, . . 𝑘𝑘 − 1 𝑖𝑖. 𝑒𝑒. 1 & 𝜎
𝑞𝑞𝑖𝑖𝑖𝑖 = −𝑙𝑙𝑃𝑃𝑙𝑙𝑝𝑝𝑖𝑖𝑖𝑖𝑝𝑝𝑖𝑖𝑖𝑖
Maximum a posteriori (MAP) parameterization is maximized through Expectation-Maximization algorithm
Markov-Switching Dynamic (abrupt-change) Probabilistic Model
Data quality checked…
Results: Two-regime constant-only Models (summary of all trips)
• Two distinct driving regimes exist – empirical data strongly favor MSDR
• Category 1 & 2 trips, drivers decelerate at higher rate• Deceleration is more volatile
Regime DurationsExpected Duration Category 1
tripsCategory2 trips
Acceleration 75 25
Deceleration 58 53
Transition ProbabilitiesCategory 1 trips
Acceleration Deceleration
Acceleration 0.918 0.082
Deceleration 0.039 0.961
Category 2 trips
Acceleration Deceleration
Acceleration 0.912 0.088
Deceleration 0.043 0.957
Short-term regime predictions •Potential to predict driving regime at specific instant of time
•Regime - stochastic event - Description of probability law governing the
change from regime 1 to regime 2 and vice versa
•Example: 25-minutes trip undertaken on freeway I-94 in Ann Arbor, MichiganSmoothed probabilities using all-sample information
Dynamics of Driving Regimes Extracted from Basic Safety Messages Transmitted Between Connected Vehicles
Short-term regime predictions •Potential to predict driving regime at specific instant of time
•Regime - stochastic event - Description of probability law governing the
change from regime 1 to regime 2 and vice versa
•Example: 25-minutes trip undertaken on freeway I-94 in Ann Arbor, MichiganSmoothed probabilities using all-sample information
Dynamics of Driving Regimes Extracted from Basic Safety Messages Transmitted Between Connected Vehicles
Acceleration associated with lower volatility
How is Driving Volatility Related to Intersection Safety in a Connected Vehicles Environment?
Kamrani, M., Wali, B. and Khattak, A. J., 2017. Can data generated by connected vehicles enhance safety? Proactive approach to intersection safety management. Transportation Research Record: Journal of the Transportation Research Board (2659) DOI: http://dx.doi.org/10.3141/2659-09.
Wali. B., Khattak, A.J., Bozdogan, H. How is Driving Volatility Related to Intersection Safety in a Connected Vehicle Environment? A Full Bayesian Perspective. Transportation Research Part C: Emerging Technologies (In-Review).
Location-Based Volatility
Motivation
• Location-based volatility >> leading indicators of crashes
• Objective: identify locations where crashes are waiting to happen
Methodological: Frequentist vs Full-Bayesian Approach, unobserved
heterogeneity
Descriptive Statistics
Traditionally, standard deviation/variance has been used to capture dispersion.
COV as a new safety surrogate…
Accidents waiting to happen (visual illustration)
Top Panel (A): Crash frequency &Intersection-specific volatilities
Bottom Panel (B): focuses on subset of intersections which are highlighted in top panel. Solid circles refer to “known hot-spots” and dashed circles refer to intersections where “crashes may be waiting to happen”
Full Bayes Hierarchical Random Parameter Models
Table: Full Bayes Gibbs Sampler Random Parameter Estimation of All Intersections
Full Bayes Hierarchical Random Parameter ModelsTable: Full Bayes Gibbs Sampler Random Parameter Estimation of All Intersections
See paper for details on segmented models for signalized intersections
Difference in Elasticity Effects Across Models
A one percent increase in volatility increases crashes by (after controlling for observed & unobserved factors):• 0.37% for all intersections • 0.66% for signalized intersections.
Note differences across Bayesian & heterogeneity-basedBayesian models
Driving Volatility, Crash Propensity, & Injury Severity
Wali. B., Khattak, A.J., Karnowski, T. Exploring Microscopic Driving Volatility in Naturalistic DrivingEnvironment Prior to Involvement in Safety Critical Events. Accident Analysis and Prevention (In-Review).
Event-Based Volatility in Naturalistic Driving
Wali. B., Khattak, A.J., Karnowski, T. The relationship between driving volatility in time to collisionand crash-injury severity in a naturalistic driving environment. To be presented at 2018 TRB Annual Meeting. Analytic Methods in Accident Research (In-Review).
On-Board Data Acquisition
Systems
Event Detail Table
Event Types:1. Safety Critical Events:
• Crash /Near-Crash2. Baseline Events: Normal
Driving
Control Factors
Large-Scale Data Analytics
Volatility Indices
Different Performance Measures
Characterize driver performance prior to
involvement in safety critical eventsComparison Group: Baseline events
Driving Volatility Concept:
Empirical toolbox:- ANOVA- Discrete Choice Framework
Key Idea:Use traditional & emergingdata sources to understandlinks b/w driving volatility &crash propensity
Unobserved heterogeneity
&omitted variable bias
Impacts:• Proactive safety
approach• Behavioral
countermeasures• Alerts &
warnings
Event volatility and crash propensity
Event volatility and injury severity
Crash severity, given a crash
Volatility measure based on entire30 seconds vehicle kinematics data
Volatility based on 1st
10-sec bin(Intentional
volatility)
Volatility based on 2nd
10-sec bin
Volatility based on 3rd
10-sec bin(Unintentional
volatility)
Event Detail Table
On-Board Data
Acquisition Systems
Knowledge-Base: Spatial Distribution of Ped-Vehicle Crashes across the US
• Analysis of current pedestrian-vehicle fatal crashes and conflicts
• Development of future travel scenarios (CAV, behavior shifts)
• Assessment of future pedestrian-vehicle conflicts
Key “Fatal” Pre-Crash Behaviors-peds (80%)• Dart-out/Dash (9.0%)• Failure to yield right of way (21.6%)• Improperly present at roadway (11.9%)
- Standing, Working, Lying, Playing• Dark clothing/Not visible (10.2%)• Inattention (1.2%)
- Talking, Eating
2018 TRB paper: Analysis of Crashes Involving Pedestrians across the United States: Implications for Connected and Automated Vehicles
Fatal crash rate (crashes/pop) distribution across US, 2013-2015
Kernel density distribution across US, 2013-2015 (N=12,217)
Knowledge-Base: Driving Volatility & Ped-Bike Safety Critical Events
www.roadsafety.unc.edu | October 12, 2017
• Analysis of SHRP-2 Naturalistic Driving Data• How driving volatility differs across baseline, ped-bike, and no ped-bike involved events.
NDS Data: Event Data Table, 9543 eventsBaseline sample = 7562Vehicle (no ped-bike) involved crashes = 1851Ped-bike involved crashes/near-misses = 74
0.99
1.97
2.24
0.82
1.66
1.79
0.00 0.50 1.00 1.50 2.00 2.50
Baseline
Vehicle Involved Crashes
Pedbike Involved Crashes
Volatility (Negative vehicularJerk: longitudinal direction)
Volatility (Positive vehicularJerk: longitudinal direction)
Key points• Greater volatility increases the likelihood of both
crash and near-crash events.• Volatility in deceleration (negative vehicular jerk),
both lateral and longitudinal, has more negative consequences than volatility in acceleration (positive vehicular jerk).
Closure… • Profound CAV impacts on mobility & safety by altering driving volatility• Data to support micro-level decisions through control assists• Customized driver feedback• Hazard anticipation & notification systems short term regime
predictions • Proactive safety trip-based, location-based, & event-based volatility
concepts• Knowledge base for vulnerable road users• Research needs/implications of our analysis
■ Predict changes in travel behavior■ Future vehicles: How will their use be substantially different?■ Network performance-partial automation-open question?