Study Type Major Data Needs Criteria/approaches to Identify Secondary Crash
Raub (1997); Karlaftis et al. (1999) static incident < clearance time+15 min(ute), < 1 m(ile)
Latoski et al. static incident < clearance time+15 min, < 3 m
Hirunyanitiwattana & Mattingly (2006) static crash data < 2 h(our), < 2 m
Zhan et al. (2008) static incident < clearance time+15 min, < 2 m, lane closure
Khattak et al. (2009) static incident < actual duration, < 1 mile upstream
Moore et al. (2004); Kopitch & Saphores (2011) static incident data < 2 h, < 2 m [both dir(ection)]
Chang & Rochon (2011) static incident data < 2 h, < 2 m; < 0.5 hour, < 0.5 mile (other dir)
Green et al. (2012) static crash data < 80 minutes, < 6,000ft; < 1,000ft (other dir)
Zhan et al. (2009) dynamic incident data maximum queuing model
Sun & Chilukuri (2005, 2010); dynamic incident data incident progression curves
Zhang & Khattak (2010) dynamic incident data deterministic queuing model
Haghani et al. (2006); Cho & Miller-Hooks (2010) dynamic incident + simulated traffic data Impact based on simulated speed contour map
Vlahogianni et al. (2010, 2011, 2012) dynamic incident + monitor + sensor data identify influential area by ASDA model
Yang et al. (2013ab) dynamic crash + sensor data spatiotemporal impact by speed contour map
Chung (2013) dynamic crash + sensor data crash impact region by speed contour map
BIG DATA FOR TRAFFIC SAFETY PERFORMANCE EVALUATION
Hong Yang, Kaan Ozbay, Ender Faruk Morgul, Kun Xie
UrbanMITS Laboratory, Department of Civil &Urban Engineering,
Center for Urban Science + Progress, New York University
Motivation Objectives
Literature Review
Improved Identification Algorithm
Spatiotemporal Impact of Secondary Crashes
Conclusions & Discussion
Traffic incidents cause delays & secondary crashes
Secondary crashes induce more operational & safety issues
Reducing the risk of secondary crashes is of high priority
Knowledge of secondary crashes is very limited
Identification of secondary crashes was a great challenge
Discuss the issues of existing approaches for secondary crash identification
Develop the virtual sensor based traffic data collection methodology
Improve the identification algorithm based on traditional sensor data
Develop an on-line scalable approach for automating identification procedure
Provide readily deployable identification tool for large-scale network use
No standard criteria to establish the fixed thresholds
The thresholds are too subjective to be reliable
Hardly reflect the actual impact of each incident
Possible Bias: Underestimate and overestimate
Weak assumption: maximum queue at clearance time
Queuing model-based: deterministic
Queuing model-based: off-line method
Depend on detailed incident information
Sensor-based: limit instrumented roads
Framework
Virtual Sensor
Potential Data Sources:
Google Map
Bing Map
MapQuest
Other maps with API
The travel time between two virtual sensors is measured and the average travel speed is interpreted.
A virtual sensor is a user created marker on map. It has coordinate information and can be accessed by API.
Data Validation
Virtual sensor output need to capture normal traffic & recurring congestion
Virtual sensor output need to capture incidents interrupt on traffic
Virtual sensor output should match traditional sensor output
Demos: virtual sensor vs detector
Primary vs. Secondary
What if a simple approach is used?
Actual impact exceeded Threshold 1
Actual impact is with Threshold 2
Challenge:
Missing information for establishing a threshold
Impact of each primary crash varies
Progression of QueueDynamic progression of queue:
Queue can exist even incident is cleared
Maximum queue can exist before and after the primary incident is cleared
Presence of Impact:
It can last over long time and space.
A reliable way is needed to account for the dynamic characteristics of impact of primary incident .
Step 1: Construct speed contour Maps (SCM)
Step 2: Develop a representative speed contour map (RSCM)
Step 3: Construct a binary speed contour Map (BSCM)
The speed measurements derived from the virtual sensors are
extracted to develop SCM
Each pair of virtual sensors will keep recording the data
Challenge:
How many virtual sensors are needed for a
given link?
Sensitivity: What are the internal algorithm
for crowd-sourcing data by those maps?
The time interval to obtain the virtual sensors
should be relative small (i.e., 5-minute)
Construct RSCM based
on historical data
collected by virtual
sensors
A sampling strategy is
used to sample the
recorded data
What if there is no special events?
Highlight the impact of special events
Challenge:
Does the facility have a
representative
condition?
The difference between
the SCM and the RSCM
highlights the incident
impact
The identification
algorithm is then needed
to technically determine
whether the secondary
crash exists or not
Challenge:
How to find the
difference?
What are the reasonable
criteria to define the
impact area?
Sub-step 4.1: Build line equation
Sub-step 4.2: Predict the coordination of special points
Sub-step 4.3: Detect the K cells that contain the short line segments
Sub-step 4.4: Check the binary speed measurement of each cell
Crash B is identified as a secondary crash of crash A
Crash C is NOT identified as a secondary crash of crash A
E LX X L L
E L
S SS T T S
T T
( , )j is t 1 1( , )j is t
( , )q qS T
1
( , )K
k q qk
V S T Q
Traditional approaches for secondary
crash identification is largely limited
by the availability of infrastructure-
based sensors.
A virtual-sensor approach is proposed
to collect the third-party open source
traffic data.
The virtual-sensor data is acceptable in
comparison with output from roadside
traffic sensors deployed by the
transportation agencies.
The automatic identification approach
is developed to capture the secondary
crashes in an efficient and reliable way.
The proposed approaches enable on-
line identification for secondary crash
on a large network even traffic
sensors are not available.
The approach is readily available in
practice.
If you have question, please contact: [email protected]
Development of an On-line Scalable Approach for Identifying Secondary Crashes