Study sites 50 unsignalized Introduction pedestrian crossings in - - PowerPoint PPT Presentation

Modelling accident frequency at unsignalized pedestrian crossings in Warsaw

30th ICTCT Workshop Olomouc, Czech Republic, 26-27 October 2017 Piotr Olszewski, Beata Osińska, Paweł Włodarek Warsaw University of Technology

• High pedestrian fatality rate in Poland: 23 person/year/mln pop.

– 8436 pedestrian accidents in 2016 – 868 pedestrians killed (28.6% of all traffic fatalities)

• Pedestrians in Warsaw:

– Constitute 60% of fatalities – Slow improvement

• Objective: to model accident frequency at pedestrian crossings

and to identify factors that affect pedestrian safety

• Part of project InDeV: “In-Depth understanding of accident

causation for Vulnerable road users”

Introduction

• P. Olszewski et al.
• No. 2

30th ICTCT Workshop, Olomouc, 26-27.10.2017

Study sites – 50 unsignalized pedestrian crossings in Warsaw

• P. Olszewski et al.
• No. 3

30th ICTCT Workshop, Olomouc, 26-27.10.2017

• Three weeks of

filming at 2 sites:

– Radzyminska – KEN

• All crossings located
• n 2x2 lane roads
• Roadways separated

by a median or a refuge island

KEN filming site

• P. Olszewski et al.
• No. 4

30th ICTCT Workshop, Olomouc, 26-27.10.2017

• 4 + 2 pedestrian accidents in 7 years

Radzymińska filming site

• P. Olszewski et al.
• No. 5

30th ICTCT Workshop, Olomouc, 26-27.10.2017

• 8 + 4 pedestrian accidents in 7 years
• Distribution of sites by the

number of accidents

• Police accident records:

59 accidents during 7 years

– 1 fatal, 14 serious – mean = 1.18 acc./site – variance = 2.51 -->

• ver-dispersion
• Problems typical for road accident dataset:

– over-dispersion (greater variability than Poisson distribution) – frequent zero observations – 23 sites (46%) had zero accidents

Accident statistics (2009-2016)

• P. Olszewski et al.
• No. 6

30th ICTCT Workshop, Olomouc, 26-27.10.2017

Accident prediction model

• P. Olszewski et al.
• No. 7

30th ICTCT Workshop, Olomouc, 26-27.10.2017

eq (1)

• Volumes of pedestrian and

vehicle traffic were counted

– Six sites: 24 h counts – Remaining sites: 3x1 hour counts: 7-8, 12-13, 16-17

• Pedestrian volume model

DPV = 5.65 P7 + 9.62 P12 + 2.1 P16 (R2 = 0.999)

• Motor traffic volume model

DTV = 3.87 T7 + 7.53 T12 + 4.17 T16 (R2 = 0.999)

Daily traffic volume estimation

• P. Olszewski et al.
• No. 8

30th ICTCT Workshop, Olomouc, 26-27.10.2017

Variable description & statistics

• P. Olszewski et al.
• No. 9

30th ICTCT Workshop, Olomouc, 26-27.10.2017

Results of model estimation - 1

• P. Olszewski et al.
• No. 10

30th ICTCT Workshop, Olomouc, 26-27.10.2017

• Eq (1) can be estimated assuming different statistical models:

Poisson, Negative Binomial, Zero-inflated - Poisson or NB

• Negative Binomial distribution is usually preferable – here it has:

– lower (=better) AIC value but – lower (=worse) pseudo R2 and lower significance of explanatory variables

• Test for over-dispersion: CT (Cameron-Trivedi) is not conclusive

– it shows that Poisson distribution cannot be rejected

• Conclusion:

– Poisson acc./year – NB acc./year

Poisson or Negative Binomial?

• P. Olszewski et al.
• No. 11

30th ICTCT Workshop, Olomouc, 26-27.10.2017

• Cumulative residuals can

be plotted, sorted by the sum of traffic volumes

• Poisson 1 model
• Negative Binomial model
• Both models perform

well – cumulative residuals are within 2

• std. error bands

CURE plots

• P. Olszewski et al.
• No. 12

30th ICTCT Workshop, Olomouc, 26-27.10.2017

Results of model estimation - 2

• P. Olszewski et al.
• No. 13

30th ICTCT Workshop, Olomouc, 26-27.10.2017

• More variables were included in the Poisson model
• Model Poisson 4 has the

best fit parameters and correct plot of cumulative residuals

• Significant variables:

DPV, DTV, HGV, HOUSE, PPEAK

• None of the other variables

were significant

• Final equation for the number of accidents per year:

Extended Poisson model

• P. Olszewski et al.
• No. 14

30th ICTCT Workshop, Olomouc, 26-27.10.2017

• The recommended accident prediction model is based on

Negative Binomial distribution:

• Risk factors affecting pedestrian safety at unsignalized

crossings: proportion of heavy vehicles, location in a housing area, less peaked pedestrian volume profile

• Models developed will help to establish a relationship

between accident frequency and conflict frequency

• Possible improvements: accounting for seasonal variability

(DTV -> AADT) and daily variability (peak ratio)

Conclusions

• P. Olszewski et al.
• No. 15

30th ICTCT Workshop, Olomouc, 26-27.10.2017

Thank you very much for your attention!

30th ICTCT Workshop, Olomouc, 26-27.10.2017