Road traffic incident risk assessment: Accident data pilot on Ring I of the Helsinki Metropolitan Area

The purpose of this project was to apply the Palm distribution to the analysis of riskiness of different traffic and road weather conditions introduced in a previous project (Innamaa et al. 2013), develop the method further, and find factors that statistically significantly affect traffic incident risk. The method was piloted using data from Ring-road I of the Helsinki Metropolitan Area. The study was based on registered accidents that occurred on Ring-road I in 2008-2012, totalling 1120. In addition to accident data, traffic data from eight automatic traffic measurement stations (inductive loops) and road weather data were also used. The basic methodological idea was to compare the traffic and weather circumstances just before an accident with the Palm probability of the same circumstances. The notion of Palm probability comes from the theory of random point processes, and means the probability distribution "seen" by a randomly selected point of the point process, i.e. the driver in this case (in contrast to the stationary probability, which is the probability distribution seen at a random time point). If each car driver had a constant stochastic intensity of causing an accident, then the accident circumstances would follow the Palm distribution. The idea of the method applied here is to assess the influence of circumstances on incidents by comparing the incident circumstance distribution with the Palm distribution of circumstances: differences between these distributions hint at effects of circumstances on accidents. The results showed that there were several specific weather conditions that were more common among drivers who were involved in an accident than among drivers in general. These conditions included air temperature from -6 degrees Celcius down, snowfall or heavy rain, limited visibility, and snowy or wet road surface. The results further showed that the probability of an accident is higher in conditions when a weather alarm is issued by the Transport Agency (the road operator) than in general. In addition, in weekday afternoon traffic (15-17 o'clock) the risk of accident was found to be 50% higher than generally. In night time traffic (2-5 o'clock) the risk was even higher. The results indicated that the traffic situation correlated poorly with accident risk. However, the results related to the traffic situation can be considered only indicative due to inaccuracies in the accident location information and sparseness of the traffic detector network. In conclusion, the findings suggest that the proposed method for identifying conditions where accident risk is elevated by comparing the traffic and weather circumstances just before the accident with the "Palm probability" of the same circumstances indeed works. Not all results were statistically significant due to some circumstances being rare. However, with the calculation of risk levels and Kullback-Leibler divergence, it was possible to assess the findings.