Spatial Analysis of Road Safety and Traffic Behaviour using High Resolution Multi-parametric Data

Abstract

The main objective of the present doctoral dissertation is the spatial analysis of harsh event frequencies in road segments using multi-parametric data, including (i) high resolution naturalistic driving and driver behavior data from smartphone sensors, (ii) microscopic road segment geometry and road network characteristic data from digital maps and (iii) high resolution traffic data. Naturalistic driving data were collected and processed with purpose-made spatial processing algorithms, performing critical functions such as derivation of additional geometrical characteristics, data merging and map-matching. The resulting spatial data-frames were then analyzed and modelled on a road segment basis. Moran's I coefficients, as well as merged and directional variograms were calculated. Spatial analyses were performed on two parallel pillars: (i) Prediction models were developed in an urban road network training area, with the intent to transfer them to a second urban road network testing area and assess their predictive performance and (ii) Causal models including road user behavior and traffic input data were calibrated in an urban arterial study area per traffic state, in order to investigate additional underlying correlations in an effort to further understand the phenomena of harsh braking and harsh acceleration frequencies. Geographically Weighted Poisson Regression (GWPR) models, Bayesian Conditional Autoregressive Prior (CAR) models and Extreme Gradient Boosting algorithms with random cross-validation (RCV XGBoost) and spatial cross-validation (SPCV XGBoost) were implemented. From the spatial analyses, numerous informative results were obtained. Spatial autocorrelation was identified in both harsh braking and harsh acceleration frequencies, and its range of influence was determined for each study area. In urban networks, certain geometrical characteristics were found to affect harsh braking frequencies per road segment: Segment length is positively correlated with harsh brakings, while gradient and neighborhood complexity are negatively correlated with them. Different geometrical characteristics were found to affect harsh acceleration frequencies per road segment: Segment length, curvature and the presence of traffic lights are positively correlated with harsh accelerations. For both harsh event types, pass count increased frequencies of both types of harsh events, while lane number and road type have more unclear circumstantial effects, depending on the utilized models. Furthermore, successful spatial predictions were conducted by averaging the results of all four methods, achieving accuracy of 87% for harsh brakings and 89% for harsh accelerations. In urban arterial segments, segment length and pass count were consistently positively correlated with harsh event occurrence overall. In addition, it was determined that different variables are significantly correlated with harsh event occurrence per traffic state: For harsh brakings in free flow conditions, speed difference between traffic and driver was found to exert a positive influence, while the influence of the averaged standardized current traffic volume was found to be negative. In synchronized flow conditions, average occupancy assumes a statistically significant positive correlation for harsh braking frequencies, while the influence of traffic volume was found to be circumstantially negative. For harsh accelerations in free flow conditions, the influence of average occupancy was found be consistently positive, as was the average mobile use seconds of drivers. In synchronized flow conditions, traffic volume was found to be positively correlated with harsh accelerations as well. In both traffic states, geometric and road network characteristic variables were found to have very circumstantial effects.