New Approaches for Traffic State Estimation: Calibrating Heterogeneous Car-Following Behavior using Vehicle Trajectory Data Dr. Xuesong Zhou & Jeffrey Taylor, Univ. of Utah 1
Outline 2 Background on Dynamic Time Warping (DTW) Application to Newells Simplified CFM Calibration Results Important Considerations
Motivations: I 3 Real-time Traffic Management Automatic Vehicle IdentificationAutomatic Vehicle LocationLoop Detector Video Image Processing
Motivation 2: Self-driving Cars as Mobile Sensor Controlled, coordinated movements Proactive approach Applications Automated cars Unmanned aerial vehicles 4
Motivation 3: Detecting Distracted/Risky Drivers 5
Underlying Theory: Cross-resolution Traffic Modeling 6 Reaction distance/spacing δ Reaction time lag τ W = δ / τ Time Space
How to Estimate Driver-specific Car-following Parameters? Input and output 7
Intro to Dynamic Time Warping (DTW) 8 Matches points by measure of similarity
Euclidean Vs Dynamic Time Warping Euclidean Distance Sequences are aligned one to one. Warped Time Axis Nonlinear alignments are possible. Eamonn Keogh Reference: Eamonn Keogh Computer Science & Engineering Department University of California - Riverside
Construct Cost Matrix for Traffic Trajectory Matching 10
Cumulative Cost Matrix 11 Dynamic programming Calculate the least cost for matching a pair of points Warp path Least cost matching points from end to beginning Singularity
Application to Newells Model 12 Follower separated by leader by reaction time and critical jam spacing Algorithm finds optimal τ n (time lag) for best velocity match Calculate d n for all time steps along the trajectory
Calibrated Parameters: Car 1737 Reaction Time Lag (sec) Critical Spacing (m) Backward Wave Speed (km/h) Avg St. Dev
NGSIM Data: I-80 Lane 4 14
NGSIM Data: I-80 Lane 4: Reaction Time Distribution 15 Mean = 1.48 seconds
NGSIM Data: I-80 Lane 4 Critical Spacing Distribution 16 Mean = 8.06 meters
NGSIM Data: I-80 Lane 4 Wave Speed Distribution 17 Mean = km/h
Current Issues in DTW Application 18 Singularities Locations with more than one match solution Data reduction algorithms Parameter estimates differ with available methods
Singularities 19
Singularity Implications 20 1 st Interpretation: Many responses to 1 stimulus 2 nd Interpretation: 1 response to many stimuli 3 rd Interpretation: Algorithm drawback Increases uncertainty in parameter estimates LCSS force 1-to-1 match LCSS : Longest Common Subsequence
Singularities Without Prior Information With Prior Information 21
Data Reduction Algorithms 22 Piecewise Linear Approximation/Regression – Somewhat subjective in application, needs dynamic parameters – Difficulties creating new points application with Newells model
Potential Applications 23 Analyze intradriver heterogeneity Markov Chain Monte Carlo method for reaction time/critical jam spacing Analyze relationships between parameters
Markov Chain Transition Matrix 24 Reaction Time Acceleration Sum1.000 Hypothetical case:
Trajectory Prediction (MCMC) 25 ~ 5% MAPE