Real-time Estimation of Accident Likelihood for Safety Enhancement Jun Oh, Ph.D., PE, PTOE Western Michigan University March 14, 2007
Background / Motivation Is it possible to predict occurrence of accidents? Maybe NOT. / Almost impossible Are there certain traffic conditions that lead to more accidents? Maybe YES. Then, is it possible to identify such traffic conditions? What will be possible indicators?
Contents Previous Studies Traffic Dynamics and Accident Empirical Example Accident Likelihood Estimation Issues on Accident Study Advanced Surveillance System
So far, previous studies... Analyzed long term historical data To identify relationships between traffic variables or geometric elements and accidents off-line studies Incident detection and incident traffic management after-incident
Objectives To enhance traffic safety under ITS To identify traffic conditions leading to more accidents Real time Before accident To estimate accident likelihood
Occurrence of Traffic Accidents Accident Traffic Dynamics Environment Vehicle Characteristics Driver Characteristics
Accident Indicator TIME Traffic Dynamics (Indicator) Normal traffic condition Disruptive traffic condition T T-x Implication starts Accident occurs
Empirical Example Freeway traffic data I-880, California Volume, Occupancy, and Speed (double-loop) 10-second periods from upstream detector stations Accident profiles (52 accidents) Traffic Variables Occupancy, Flow, and Speed 5 minute Mean and STD
Pattern Classification Two traffic conditions Normal traffic condition: a 5-minute period apart from traffic accident (more than 30 minutes apart) Disruptive traffic condition: a 5-minute period right before an accident Non-parametric density estimation kernel smoothing technique Best indicator: STD of speed
Estimation PDF
Bayesian Model for Accident Likelihood P(A/X) = Posterior probability that given traffic measurement belongs to traffic conditions leading to an accident occurrence P(A) = Prior probability that given traffic measurement belongs to disruptive traffic condition P(N) = Prior probability that given traffic measurement belongs to normal traffic conditions
Estimation of Accident Likelihood
Real-time Application
Identification of Accidents * The percentage of time when P(A/X) was above the given threshold Threshold # accidents identified % accidents identified % time*
GIS Database for Enhanced System Traffic Accident Data Mapping Linear Referencing & Dynamic Segmentation Reconstruction of highway segments Detector location and accident location Other Characteristics Weather Highway Geometry Real-time Traffic Data
Database Example Real-time Traffic Data Accident location and type
Possible Application Framework Real-time estimation of accident likelihood Provide safety information at upstream via VMS Real-time traffic measurement with highway geometry and weather Drive safely Caution! Traffic Unstable Is traffic condition stable? No Yes
Issues on Accident Study Accident data availability and accuracy Need more data Accurate accident occurrence time Accident duration Other measures? Wide-area detection Individual vehicle tracking Need better surveillance systems
An Advanced Surveillance System Present traffic surveillance systems mostly use inductive loop detectors (ILDs) have significant limitations (e.g. point estimates) and errors reduce the ability to effectively manage and control freeway and arterial traffic systems, and to implement ATMIS Advanced sensor systems Integration of weather and surface sensors Individual vehicle detection for details Vehicle reidentification techniques utilizing existing and future infrastructure
Vehicle Reidentification Matching Inductive Vehicle Signatures Volume (veh/interval) Travel Time (sec) Downstream OfframpLane 1 HOV Lane 2 HOV Lane 3Lane 4Lane 5Lane 6 Up-stream Up-stream Lane 1 HOV-81 veh sec 18 veh sec 6 veh sec --- Lane 2--3 veh sec 203 veh sec 4 veh sec 5 veh sec - Lane veh sec 164 veh sec 16 veh sec - Lane veh sec 108 veh sec 2 veh sec Lane 527 veh sec ---4 veh sec 21 veh sec 69 veh sec Lane 610 veh sec veh sec 16 veh sec Lane veh sec 5 veh sec Volume Occupancy Speed Vehicle Types Section Density Section Delay Travel Time Level of service Lane-by-lane travel time Lane changing pattern
Concluding Comments Speed variance can be a good surrogate Traffic dynamics reflects hazardous factors Temporal spatial speed variation Advanced surveillance systems may provide better exposure Lane-by-lane travel time Lane-changing pattern Possible to identify traffic conditions leading to more accidents (Accident Likelihood) Integration of traffic, weather, and geometry information
Thank you Q & A Jun Oh