ROAD ACCIDENT ESTIMATION MODEL IN URBAN AREAS

Summary. Urban areas are significantly different in terms of traffic risk. In a decisive manner, they are the result of...


INTRODUCTION
From a phenomenological perspective, the urban area where traffic risk occurs is a complex system. Four main classes of inputs could be considered significant for risk circumstances [1]: 1) Objects class-defines the categories of elements, such as mobile entities (vehicles, cyclists, pedestrians, etc.) or features of technical and road urban infrastructure, which intervene in risk situations; 2) Actors class-defines different groups or organizations which could influence the system: urban authorities, urban network management companies, civil associations, population; 3) Spatial structure class-describes the topological configuration of the different elements and phenomena related to risk; The interactions between these classes generate a traffic flow characterized by the uncertainties related to traffic risk. In the last two decades, all four classes have suffered drastic transformations in Bucharest, the study area of the research presented in this paper. The rapid and radical changes of socioeconomic life of the Bucharest population, major changes in urban commerce structure caused by the emergence of large commercial centres, new concentration of urban and suburban residential areas, spatial and structural changes of places of interest (for work, education, leisure, etc.) have led to changes in the size and pattern of urban traffic flow. Consequently, the traffic risk exposure has increased. Understanding the road accident causes, identifying appropriate solutions and applying them since the urban planning phase can lead to traffic safety improvement [2,3]. In this frame, the aim of our research is to identify the relationships between land use and road safety performances and to develop a road accident estimation model at urban zones level.
Several models are developed to estimate the number of accidents [4]. Most of the developed models include accident prediction function for different classes of junctions and road segments [5 -11]. Reduced number of models have been elaborated for emphasizing the influence of the urban zone characteristics on accident occurrence. There are models developed for identifying the relationships between road accidents and compound road environments [12], for street network [13] and for land use [2,14]. This paper presents the case study of the model developed to estimate accident function of the urban area attributes in Bucharest. Due to the peculiarities of Bucharest urban zones and the constraints given by the available data, new examination procedures have been necessary. The initial analysis has been made at the land use zone level, but unsatisfactory statistical measures resulted. Consequently, the analysis level has to be changed and adequate statistical measures have been obtained for the accident estimation model developed at the traffic analysis zones level.

STUDY AREA AND MODELLING SETTINGS
Our analysis of the relationships between road accidents and land use in Bucharest started from the following input data: -Urban road network geo-database, which includes attributes of junction and road sections; -Statistics of road accidents recorded in Bucharest during 2008 -2012; the records offered by the Bucharest Road Police Department, including severe road accident with injuries and/or significant financial damages ( Fig. 1), were edited in the geo-database and the data set with accident location on urban road network was obtained (Fig. 2). The GIS procedures were applied to obtain classes of different accident types (vehicle-vehicle accidents and vehiclepedestrian accidents) on categories of road network features (junctions and road sections). The sample dimension of each accident class was too small to allow proper calibration of the estimation model [15]. Consequently, we have processed and calibrated the model for the aggregate set of all severe accidents recorded in Bucharest. -Land use zones geo-database, whose available data includes the area of Bucharest city (around 228 sq. km with 1,890 mill. inhabitants) divided in N LU = 9675 zones for 12 land uses (Tab. 1). Because our aim is to identify the influence of land use on traffic risk, we eliminated from the land use dataset the areas of road infrastructure. Even if the accidents are located on road network features, the influence of land uses of the areas located around the accident location is analysed. After this step N LU * = 7650 land use zones are selected for further analysis and the recorded accidents are assigned to adjacent land use areas.
The land use and accident date sets are processed in the GIS environment to obtain the accident location distribution in urban zones. The total number of accidents A i assigned to zone Z i , i=1..N LU * is calculated by the equation:

35
where k I A is the number of accidents that occurred at junction I k located in zone Z i and XY A is the number accidents that occurred at section XY located in zone Z i . The results (Fig. 3) show that a significant rate of accidents is located in areas with a high density of buildings (88.71%), followed by areas with a low density of buildings (4.24%), educational use (3.33%) and commercial zones (1.69%).
The land use of each urban zone has a direct influence on the generated and attracted traffic flow and also on the activities and socioeconomic characteristics of the neighbouring zones. Therefore, it is necessary to evaluate the spatial autocorrelation of accident distribution before analysing the relationships between land use and traffic risk. The spatial autocorrelation was assessed using Moran's I Index [16,17] computed function of the number of accidents assigned to urban zones. We obtained the value of 0.39 for Moran's I Index with a z-score of 264, which indicates a clustered pattern of accidents and spatial relationships between accidents located in adjacent zones. Hence, the accident analysis at this level of urban zonation is not suitable and we had to choose another level of analysis. We decided to analyse the accident distribution in the traffic analysis zone (TAZ) level for N Z =80 zones (Fig. 4) defined by the Bucharest General Master Plan of Transport [18]. We applied the previous procedure (eq. 1) to assign accidents to the analysed zones and we again evaluated the spatial autocorrelation of the accident distribution on the new level of zoning. In this case we obtained the values 0.009 for Moran's I Index (very low) with a z-score of 0.68 and a p-value of 0.49, which indicate a random accidents distribution and the fact that the accidents located in a zone do not influence the accidents in adjacent zones. Therefore, the obtained data sets are appropriate to assess the relationships between urban zone characteristics and traffic risk.

MODEL DEVELOPMENT
GIS spatial analysis procedures are applied on datasets of TAZs (including accident distribution), urban road network and land use to obtain features and attributes corresponding to accident analysis on selected urban zones (Fig. 5). The resulting data set allowed for the calculation of the variables Y i , i=1..12 (Tab. 2) on the analysis zones used to estimate the relationships between urban characteristics and road accidents (the number of accidents represents the dependent variable, denoted by D v ). We considered the sum of the accidents on urban zones during 5 years in order to use a more representative data sample. Table 2 summarizes the minimum, maximum and mean values of urban characteristics for N Z = 80 zones.  After several tests of correlation applied to different subsets of two up to five variables of the set {Y i │i=1..12} using correlation coefficient R 2 , adjusted correlation coefficient R 2 , Akaike information criterion AICc and variance inflation factor VIF as evaluators [16,17], we identified two classes of significant variables in road accidents:

Accident Analysis Zone Data Set
I. Population of high-density building development zone (P HDU ), area of road infrastructure (S TU ), area of educational zones, (S EDU ) and area of commercial zones (S COM ); based on the ratio between average and variance of these variables [3,19] used in the following equations to accident estimation: parameters to be calibrated.
with Z N j ..
parameters to be calibrated.
Using GIS spatial regression analysis tools, we have tested the estimated results based on equations (2) - (7). Better assessments were obtained using the second class of variables (densities of characteristics of urban zones) than using the first class of variables describing the total characteristics (P HDU , S TU , S EDU , S COM ).
After many statistical tests and steps of calibration, we obtained the accident estimation equation: Applying eq. (8) on the available data, the values of 0.76 for correlation coefficient R 2 and 0.71 for adjusted correlation coefficient R 2 resulted, meaning that the model can be used to estimate accidents in the study area. Traffic accidents are composite events that can be caused by a variety of factors related to intrinsic features of the road network and urban environment but also to the behaviour of traffic infrastructure users, vehicle state, time and ambient conditions. Given that the proposed model aims to determine the influence of the urban patterns and land use on accidents, we consider that the values obtained for the correlation coefficients are satisfactory and, consequently, the identified independent variables have to be considered in urban planning and in traffic risk analysis.
The outcomes of eq. (8) lead to standard residual errors less than 1.5 for 70 zones (covering around 80% of the study area) and less than 0.5 for 42 zones (around 40% of the study area). Standard residual errors over 1.5 are obtained in zones with mixed use (residential and industrial) and land use changes in the last decade (zones which include business park development, large commercial areas). Additional variables (e.g. employees per activity categories and capacity of commercial areas) are necessary for enhanced calibration, but in this phase eq. (8) can be used to estimate accidents in the assessment of alternatives to urban planning.

CONCLUSION
Land use in each urban zone (residential, commercial, offices, mixed-use areas) is decisive for generated and attracted traffic flows. A new urban development has a direct influence on economic and social activities and implicitly on traffic flow and its associated risks. Thus, the model presented in this paper aims to identify the correlation between land use and traffic risk.
The decisions on urban planning (development of new residential areas, building of new business parks, location of large commercial centres) need rigorous validation of the consequences for road traffic and the associated accident risk. The size of traffic risk has to intervene as an additional criterion in these types of decisions.
The model presented in this paper represents a useful tool for assessments of the consequences of land use on road accidents. Using GIS procedures and spatial analysis models, the density of population of HDU zones and the density of road infrastructure area are identified as the most significant variables in road accident occurrence in the Bucharest city area. The density of educational zones and the density of commercial zones are also included as variables in the accident estimation model.
Better calibration could be obtained if the density of educational zones and density of commercial zones are computed as functions of their capacities instead of their surfaces. Furthermore, because HDU zones also include business parks, better estimation could be found if, beside the density of the population of HDU, the density of employees in categories and in urban zones will be analysed. Nevertheless, until a more detailed database is gathered, the developed model of the available data can provide significant results (with 0.76 for correlation coefficient R 2 ) on accident estimation function under urban zone characteristics and represent an important step in the development of a road accident estimation model for the Bucharest area.