In short Identifying hazardous locations at intersections by - - PDF document

Identifying hazardous locations at intersections by automatic traffic surveillance Hagen Saul, Marek Junghans and Andreas Leich German Aerospace Center (DLR) ICTCT 2015, 29.-31.10.2015, Ashdod, Israel

• Measurement campaign at known hot spot
• Observe traffic, conflicts with cyclists involved
• Are conflicts (and which conflicts) detected automatically?
• Can they be clustered, thus indicating hot spots?

In short

• Motivation
• Data acquisition / measurement campaign
• Results of analysis
• Conclusion
• Outlook

Overview

Automatic traffic surveillance: 1.provides densely sampled data in time and space

• 1. trajectories (traffic safety)
• 2. Also: traffic volumes, velocities etc.

2.Allows to collect all incidents (unreported ones!) 3.Pre-selection of critical incidents 4.Enables pre-conflict analysis 5.Enables long-term evaluation of traffic safety without accidents (even at locations with few accidents!) 6.Cheap, once established But requires sufficient accuracy…

Motivation—in general

• Goal of the measurement campaign is test our system (fully automatic) and

continue the research regarding traffic safety of cyclists.

• Can tendencies/correlations be shown or revealed?

Motivation—for field study

• Provided by police Berlin for 2014:
• crashes with cyclists involved: 7699 (6952 in 2013) – rise of 10%
• share of crashes with cyclists involved: 5.8%
• 2001: 4.06%, 2002: 4.37%, …, 2013: 5.31%)

(number of total crashes , all modalities , had been increasing since 2006!)

• On average 21 crashes a day
• Vulnerable: every 4th fatality is a cyclist
• Main cause (not bicycle): wrong behavior when turning

Official statistics

• Number of crashs with involvement of cyclist in 2014 in dependence of

daytime

Current situation

• 15 crashes recorded at Prinzenstraße/Moritzplatz in 2014
• Conducted on July 10th 2014, 6am-6pm

Measurement campaign

• MUSE vorstellen

• Trajectory-based analysis
• Virtual / optical loops

Analysis—cyclist trajectories

8-9am 6am – 6pm

• Traffic intensities
• f cyclists (groundtruth)

Analysis—Traffic Parameters

• Velocity of cyclists in free flow (6am-7am)
• Peak: 5-6 m/s² (18-22 km/h)

Analysis—Velocities

• Velocity of cyclists at max. traffic intensity (14pm-15pm)
• Peak: 4-5 m/s² (14-18 km/h)

Analysis—Velocities

• if collision predicted , 1s forecast period
• 252 conflicts in total

Conflicts

• if collision predicted , 1s forecast period
• 252 conflicts in total

Conflicts

• 2-3pm

TTC

TTC

DRAC

DRAC

• 2-3pm

DRAC

• Determined positions of max. deceleration for every cyclist
• Originally: try to determine braking entry point, in order to get a clue, when

possible hazard is recognized and reaction (braking) is triggered

• Also possible: point of max. jerk when decelerating (earlier than point of
• max. deceleration)
• Further research necessary, no knowledge, if and how good this entry

point can be determined from outside (i.e. by camera)

Braking

• Tbd ViewCar-speed and velocity during braking maneuver

Braking

• 11-12am

Braking

Braking

Incidents

Incidents - 1

Incident 1: trajectories recorded

Incidents - 2

• Three hot-spots identified
• Maybe stronger correlation traffic volume of vehicles leaving roundabout to

number of conflicts with cyclists involved

Summary and Conclusions

• There are indications for hot-spots
• Hot spots are not indicated at any aggregation level (here: 1h)
• Only indicator, because errors of object detection and error propagation for

derived parameters (e.g. TTC)

• And data can be unrepresentative
• Validity only by long-term analysis (effects depending on day of week, daytime,

month)

Summary and Conclusions

• 1. Improve image processing (improve detection rates and accuracy

trajectories)

• Reduce false positive incident detections
• Already very good results for less complex scenes
• Fusion of several cameras
• 2. Are locations of clusters of conflicts in accordance to (possible) clusters of

crashs (crash data, geo-referenced, police Berlin)?

• 3. Further analysis of hazardous locations:
• Is (gravity point of a) hot spot stable over time (what times)?
• 4. Determine braking entry points for strong braking maneuvers

Outlook

• Comments / criticism / improvement suggestions ?

Hagen Saul German Aerospace Center Institute of Transportation Systems Rutherfordstraße 2, 12489 Berlin hagen.saul@dlr.de

Thank you!

Accuracy object tracking

Accuracy object tracking

0m 5m 10m Kennwerte Eigene Entwicklung

Mittelwert | Standardabweichung | Median

2,36 m | 2,46 m | 1,70 m

Accuracy object tracking

• A Comparison of Trajectories and Vehicle Dynamics Acquired by High

Precision GPS and Contemporary Methods of Digital Image Processing.

• A Comparison of Methods for Detecting atypical Trajectories.
• Road user tracker based on robust regression with GNC and preconditioning.

In: Video Surveillance and Transportation Imaging Applications

• Calculation of Error Rates for the Detection of Critical Situations in Road Traffic

Further research using MUSE