MICROSCOPIC ROAD SAFETY COMPARISON BETWEEN CANADIAN AND SWEDISH ROUNDABOUT DRIVER BEHAVIOUR Canadian Association of Road Safety Professionals Conference 2017 Paul St-Aubin, Ph.D. 1,2,3, Nicolas Saunier, Ph.D. 2, Luis Miranda-Moreno, Ph.D. 3, and Aliaksei Laureshyn, Ph.D. 4 1 2 3 4 1 / 27
RATIONALE Road safety varies considerably throughout the world Death rates from road traffic accidents by country, per 100,000 inhabitants, world map (WHO 2012) https://en.wikipedia.org/wiki/list_of_countries_by_traffic-related_death_rate 2 / 27
RATIONALE Yet, even if we compare countries of broadly similar makeup, such as Canada and Sweden, which have comparatively similar population densities, levels of urbanization, Climates, levels of economic development, and levels of social development, we still find important differences between the two. 3 / 27
RATIONALE Overall, Canada has roughly twice as many accidents as Sweden in a number of different metrics. Statistic Canada Sweden Ratio Year Source Fatalities per 100,000 inhabitants 6.8 3 2.26 2010 WHO 2013 Fatalities per 100,000 inhabitants 5.5 2.7 2.03 2013 OECD 2015 Fatalities per 10,000 registered motor vehicles Fatalities per billion veh-km travelled 0.85 0.45 1.88 2013 OECD 2015 5.6 3.4 1.65 2013 OECD 2015 Accidents per 100,000 inhabitants 501 210 2.39 2013 OECD 2015 4 / 27
RATIONALE Some important differences exist in road design philosophy which may explain the discrepancy. A notable example is the universal use of stop signs in North America versus yield signs or no traffic control at all at many European intersections. 5 / 27
OBJECTIVE Surely road design and urban makeup can t explain it all. To what degree is this discrepancy in road safety explained by differences in the drivers themselves? There are many possible local factors that may affect driver conditioning. The focus of this work is on studying the observable differences in driver behaviour in great detail using novel traffic data collection techniques. While capturing the individual factors which may explain driver conditioning is not a priority for this study, it is important to keep these potential source in mind for future work. 6 / 27
OBJECTIVE Roundabouts are selected for comparison between Sweden and Canada: Design is highly similar in both jurisdictions Modern roundabout invented in Europe and adopted in North America with minimal design changes by comparison to other types of road infrastructure Rules of priority are handled implicitly and entirely by the drivers. Many opportunities for negotiating, conflicting manoeuvres, interactions, and etiquette to manifest itself. 7 / 27
MERGE ZONE A number of roundabouts are included in the study, but each is deconstructed into it s fundamental merge zones, forming the basic unit of analysis. 8 / 27
SITE SELECTION 10 merging zones (at 4 roundabouts) in Sweden and 9 merging zones (at 7 roundabouts) in Quebec, Canada. Identical speed limits Comparable land use and urban density Comparable geometric dimensions, number of lanes, layout, etc. At each merging zone, the following factors are parametrised and included in the model: Land use, urban density, outside radius, hourly flow, flow ratio, years since built Accident data available, but considered too unreliable. 9 / 27
Site Urban Outside Hourly Flow Land Use Flow Ratio Construction (Sweden) Density Radius (m) (veh/h/ln) Year 1-1 Mixed Medium 25 408.4-0.432 1965 4.1 1-2 Mixed Medium 25 394.7 0.283 1965 4.1 2-1 Mixed Verylow 22 281.8 0.293 2003 1 2-2 Mixed Verylow 22 289 0.517 2003 1 2-3 Mixed Verylow 22 226.8 0.252 2003 1 2-4 Mixed Verylow 22 218.4 0.934 2003 1 3-1 Mixed Low 22 123.9 0.646 2010 1.5 3-2 Mixed Low 22 121.4 0.568 2010 1.5 4-1 Residential Low 16.5 191-0.417 1995 1.4 4-2 Residential Low 16.5 142.9 0.054 1995 1.4 MEAN 21.5 239.9 0.27 1995 1.8 Accidents per Year Site Urban Outside Hourly Flow Land Use Flow Ratio Construction (Québec) Density Radius (m) (veh/h/ln) Year 1 Residential Medium 25 315.1-0.327 2004 7 2 Residential Low 22.5 178.8 0.421 2003 1.4 3 Mixed Low 24.5 51.5 0.6 2003 7 4 Residential Low 18.5 236.9-0.361 2003 0.7 5-1 Residential Low 15.5 64.6-0.518 2004-5-2 Residential Low 15.5 93.4 0.607 2004-6 Residential Low 18.5 46.6 0.112 2005 1 7-1 Mixed Medium 18 150.8 0.608 2004 7.7 7-2 Mixed Medium 18 238.6 0.534 2004 7.7 MEAN 19.5 152.3 0.186 2004 4.19 Accidents per Year 10 / 27
TRAFFIC DATA COLLECTION To capture driver behaviour in great detail, a largescale, video-based automated traffic data collection system is employed. Large quantities of traffic data can be collected costeffectively: all-day monitoring across multiple sites. Video captures all road users continuously, at a high resolution (up to 30 times per second). Video tracking is less prone to bias and boredom, and processing can be rerun to refine result as the technology matures. 11 / 27
TRAFFIC DATA COLLECTION Traffic Intelligence Application of computer vision tvalib Traffic analysis and DB mgmt. 12 / 27
TRAFFIC DATA COLLECTION 13 / 27
MEASURES OF BEHAVIOUR Speed: Simple to conceptualise and measure (in km/h) Widely used throughout the literature Excellent predictor of collision severity Yielding Post-Encroachment Time (ypet): Simple to conceptualise and measure (in seconds) Essentially an accepted gap time measured between leading and lagging road users when an approaching road user enters the merging zone Possible predictor of collision probability 14 / 27
MEASURES OF BEHAVIOUR ypet lead ypet lag 15 / 27
MEASURES OF BEHAVIOUR Time-to-collision (TTC): Advanced surrogate safety measure (measured in seconds) Models potential collision courses ( conflicts ) continuously and evaluates the time remaining until the collision would occur If user have, overall, more time to react to these collision courses, they are more likely to be able to avoid them Possible predictor of collision probability 16 / 27
MEASURES OF BEHAVIOUR The TTC model used is the Discretized Motion-Pattern motion prediction model (DMP): Purpose-built general model capable of modeling complex, non-linear driving environments, such as with roundabouts Machine learning teaches the model to predict collisions using existing road user movement at each individual site. 17 / 27
MEASURES OF BEHAVIOUR 18 / 27
ADDITIONAL FACTORS CONSIDERED Individual interaction properties include: 15-second exposure: number of road users present in the scene 7.5 seconds before and after an interaction Interaction angle: angle between the heading of both road users at the instant of interaction Speed of the slower road user Speed of the faster road user These are relevant for TTC observations and are measured alongside TTC and any collision courses modeled. 19 / 27
RESULTS - SPEED 20 / 27
RESULTS - SPEED Speed profiles Québec Sweden 21 / 27
RESULTS - SPEED Linear Regression of Mean Speed Speed significantly lower at Swedish sites Flow also significantly associated with speed (unsurprisingly) Other factors tested not significant Recall: speed limits identical at all sites Coefficient P > t constant 35.90 0.000 Swedish Site -4.460 0.007 Flow (per hour per lane) -0.024 0.003 Adjusted R 2 0.658 Wald prob. > F 0.000 Observations 19 22 / 27
RESULTS - ypet Lead ypet cannot be explained by any factors. Linear Regression of Median Lag ypet: Swedish sites not observed to have significant correlation Lag ypet increases with outside radius Lag ypet decreases with increasing flow Recall: smaller ypet generally suggests more aggressive behaviour Coefficient P > t constant 1.303 0.018 Outside Radius (m) 0.0798 0.010 Flow (per hour per lane) -0.003 0.001 Adjusted R 2 0.447 Wald prob. > F 0.003 Observations 19 23 / 27
RESULTS - TTC Random effects regression of individual TTC observations: Increased TTC at Swedish sites Increasing TTC with 15-Second micro exposure and Exposure interaction angle Recall: smaller TTC generally suggests more likely collision Coefficient P > t constant 0.583 0.000 Swedish Site 0.293 0.029 0.0169 0.000 Interaction Angle 0.0032 0.000 Within R 2 0.054 Between R 2 0.424 Overall R 2 0.020 Wald prob. > F 0.000 Observations 23565 Groups 19 24 / 27
CONCLUSIONS Driver behaviour is observed to be safer at Swedish roundabouts than at similar Québec roundabouts, even after controlling for land use, urban density, posted speed limits and signage, geometry, layout, number of lanes, traffic characteristics, and construction year. This trend is consistent across three measures of behaviour serving as surrogate safety measures: Lower speeds (less severity in the event of a collision), Higher post-encroachment times, and Higher time-to-collisions (both leading to more likelihood of avoiding potential collisions). 25 / 27
CONCLUSIONS This trend may explain the observed disparity in road safety records between the two countries, and is furthermore consistent with the accident history recorded at each roundabout (though the influence of individual merging zones and individual users is unknown given the lack of resolution of the accident records) Future work might involve more thorough comparison with finer accident records (if available), and additional investigation into individual factors that could explain latent differences in behaviour. 26 / 27
Thank you! This research made possible by the Fonds de recherche du Québec Nature et technologies, the Ministère des Transports du Québec, the Fonds de recherche du Québec Santé (proposal number 2012-SO-163493), and the Réseau de Recherche en Sécurité routière. 27 / 27