Pedestrian Accident Prediction Model for Rural Road

Pedestrian Accident Prediction Model for Rural Road A.K.Sharma Civil Engineering Department RCOEM Nagpur, India sharmaakrkn@yahoo.co.in V.S.Landge Civil Engineering Department VNIT Nagpur, India vslandge@rediffmail.com Abstract Pedestrians, one of the Vulnerable Road User, have become more susceptible to traffic crashes with the rapid growth of motor vehicles in India. In terms of pedestrian crashes on a worldwide scale over 4,00,000 pedestrians are killed every year and over 10,000 pedestrians are killed on Indian roads. To date, only limited research has been undertaken to develop the accident prediction model for pedestrian accidents. This paper focus on pedestrian crash prediction model on Indian Rural Highway(NH-6). Accident data collected between 2005-09 over a stretch of 100 km of road length are used for modeling. The Negative Binomial method is used to model the frequency of accident occurrence. The Akaike Information Criterion (AIC) is used to measure the relative goodness of fit. The candidate set of explanatory variables are:, Total Traffic volume (AADT), Lane width (LW), Shoulder width (SW), and access density (AD). It is observed that access density, Shoulder width and Lane width have significant impact on pedestrian safety. Keywords-pedestrian; crash rate ; negative binomial; acesss density; shoulder width; lane width 1. INTRODUCTION Accidents today are among the leading causes of death. Motor vehicles claim the largest toll of life and tend to be more serious. Rapid urbanization, modernization and industrialization have increased the problems of traffic. In India traffic on rural highways are heterogeneous and includes fast moving vehicles like cars, buses, trucks, scooters and slow moving traffic members like bicycle, cart, cattle, and pedestrians. Around 1, 34000 persons were killed in India in 2010[1] while travelling by various modes of transport on roads. Percentage share of different modes of transport for these fatal accidents are shown in Table.1 below. The numbers of pedestrians killed were 12,204(9.1%). In terms of pedestrian crashes on a worldwide scale, over 4,00,000 pedestrian die every year in road accidents. Walking is a basic human activity and everyone is a pedestrian at one time or another. As automobile transportation continues to increase around the world, Table.1 Road Accident Deaths by Type of Vehicles (Percentage Share), (India) 2010 pedestrians have become more susceptible to traffic crashes, especially in countries like India where traffic laws are poorly enforced. Even though pedestrians are legitimate road users, they are frequently overlooked in the quest to built more sophisticated transportation systems. Most planning for better and wider roads show little concern for pedestrians. Roads are widened to accommodate larger flow of motorized traffic, little realizing that there are no pavements for pedestrians as a consequences. Many of the pedestrian accidents can be prevented if they are given right of way on pavements. This paper talks about the various factors which can influence pedestrian safety, and quantifying the impact of these factors to decide suitable measures that are required for increased safety. II. OBJECTIVE The specific objectives of studies are: CAR 9.1 THREE WHEELER 5.4 TEMPO/VAN 6.3 TRUCK/LORRY 20 BUS 9.7 JEEP 7.5 TWO WHEELER 21.1 PEDESTRIAN 9.1 BYCYCLE 2.4 OTHERS 9.4 Development of Correlation between road accidents and geometric design parameters of highway along with traffic operating characteristics for pedestrian accident. 66

Evolving engineering remedial measures for improving safety on the selected stretch. Practical recommendations for improving traffic safety on the said highway. III. STUDY AREA AND SCOPE The study area chosen is National Highway No.6 commonly refer to as NH-6 or G.E. Road (Great Eastern Road),which is a one of the most busy national highways in India. It s a connecting corridor to major states of India namely Gujarat, Maharashtra, Chhatisgarh, Orissa, Jharkhand and West Bengal. The Highway Passes through the cities of Surat, Dhule, Nagpur, Raipur, Sambalpur,Kolkata. NH-6 eventually will be one of the important links of Asian Highway network (AH-46). The scope of present study is limited to road section passing through central Indian states of Maharashtra. Maharashtra, one of the most advanced state of India has high traffic accident rate. Most of these accidents occur on the National highways. The highway under consideration has very high rate of accidents. Many of these accidents are fatal and involvements of heavy vehicles in such accidents are in large proportion. For the present study the geometric parameters like, Shoulder width (SW), Lane Width (LW) and access density (AD) and traffic parameters like, Total Traffic Volume(AADT) are taken into consideration. IV. METHODOLOGY Methodology adopted for the study is as specified below: Identification of study area Crash data collection from the law enforcement agency and insurance companies. Road geometric and Traffic parameter data from field studies Selection of variables and modeling methods Development of accident prediction models Testing of models, interpretation of results and remedial measures suggestions. Diagram showing methodology adapted is given in fig.1 V. DATA COLLECTION AND ANALYSIS Accident data was collected from the police stations and insurance companies. Road geometry and traffic data was collected through field studies and traffic count survey for a road length of 100km between Amravati City and Nagpur City of Maharashtra State in India. For the purpose of collecting road geometry data, the road was divided into segments of similar characteristics ranging from 0.2 to 0.6km length for curve portion and 1.0 to 2.2Km for straight portion. National Highway No 6 experiences the crash rate as high as 1.62 accidents per year per km.it has a very high rate of fatality 0.38 death /km/year.the highway share heavy vehicles, passenger cars, two wheelers, animal drawn carts, cattle, and pedestrians. Heavy vehicles are involved in 78% of the accidents, passenger cars are involved in 48% of accidents, two wheelers are involved in 62% of accidents and pedestrians are involved in 21% of accidents. The preliminary analysis of the data are given from fig.2 to fig.5. Fig.2 shows a positive relation between dependent variable and access density at most of the locations. It suggests that more access points to the main road will give rise to more traffic accidents. Fig.3 which gives relation of AADT with the dependent variable suggests that accident rates are more when AADT variation is between 8000 to 11500. Fig.5 gives a relation of shoulder width with accident rate, which suggest that more deficient shoulders gives rise to more accidents. Similarly relationships of other variables are presented graphically in different figures. VI. VARIABLES USED IN MODELS A. Variables Related with Geometry A thorough scrutiny and the preliminary analysis of the data led to selection of following highly influential independent parameters [2] [3]. (i) Shoulder width: Shoulder provides an area along the Highway for vehicle to stop, particularly during emergency. Slow moving vehicles, pedestrians can use the shoulder and keep the carriageway free for heavy and fast moving vehicles. (ii) Access Density: In urban and suburban areas, the rapid growth of the local economy has steadily increased the demand for access points along multilane highways. The availability of access is necessary to commercial or residential developments, usually at the expense of traffic operations and the safety of local highway systems. (iii) Lane Width: : Traffic flow tends to be restricted when lane width reduces. This is because vehicles have to travel closer together in lateral direction. Lane width available hence is treated as an important parameter. It has been found that accident rates reduce as lane width increases. B. Variables Related with Traffic characteristics (iv) Traffic volume: Traffic volume is believed to have considerable impact on the crash rate. For this study Annual average daily traffic (AADT) is used as a parameter to indicate traffic volume.traffic volume for various segments is collected and incorporated at appropriately in the model [4]. 67

VII. MODELING Various modeling techniques have been tried to model accidents aiming for accuracy. However the suitability of the model depends on the data quality and is location specific. The model building methodology is selected based on the availability of data and accuracy of the data. Various modeling techniques popularly employed are as discussed below : Deterministic models which are widely used,are not considered to be suitable for an arbitrary and sporadic event like traffic crashes. Much of the early work in the empirical analysis of accident data was done with the use of multiple linear regression models. As the literature has repeatedly pointed out, these models suffer from several methodological limitations and practical inconsistencies in the case of accident modeling. To overcome these limitations, researchers turned to stochastic models. A. Stochastic models: Stochastic models that are a reasonable alternative for events that occur randomly and independently over time. Unlike deterministic models stochastic models assume accident as random event, early in 1989; Okamoto et al. [5] suggested that the occurrence of traffic crashes follows stochastic distribution. In 1990, Garber et al.[6]. developed several models to describe the occurrence of crashes in using stochastic modeling techniques, like Poisson regression(pr) and negative binomial regression(nbr). Various studies further examined the goodness-of-fit of these regression models. More stochastic models were also proposed other than Poisson regression models and negative binomial regression model, which included Zero Inflated Poisson Regression Model and Zero Inflated Negative Binomial regression models. This paper presents models build using Negative binomial regression. B. Modeling Method: Crash being non negative, sporadic and discrete, use of deterministic models weakens the analysis. Stochastic modeling methods [7]overcome this limitation and are a better option for random events like accidents.negative binomial model was selected to model accidents. Fig.1 Diagram showing methodology adapted 68

Fig.2 Access Density vs crasht Rate Fig.3 AADT vs Crash Rate Fig.4 Lane Width vs Crash Rate 69

C. Negative binomial regression models: One of the problems that plagues Poisson regression model is over dispersion. Negative binomial regression was developed to be an improvement on the Poisson regression process. The negative binomial regression model allows for over dispersion in the model and can be used to quantify various parameters more effectively.in Miaou s[8] study, the negative binomial regression model was of the form; p(y=y i ) = y i = 0,1, 2, 3...(1) where 1 1 y i y 1 a i i 1 1 1 y 1 i i i p(y=y i ) = the probability of y accidents occurring at a given site, k x ij j j 1 i = E(Y i )= v i e And Var(Y i ) = i +α i 2 i=1,2,3,,n --------(2) Where Γ(.)= Gamma function; = rate of over dispersion. Road segment was divided into curved sections and straight section as collected data showed a vast difference in values on the curved and straight sections. Different sets of models were developed for straight and curved segments. This paper presents the models developed for the straight sections. Total accident per km per year is used as the dependent variable. Table 2 shows the variable selected for modeling. D. Model Selection Criteria: Maximum likelihood estimation method has been employed widely in estimating Poisson, negative binomial [9] and zero inflated regression models. According to definition of maximum likelihood estimation method the estimated parameters are best when the maximum value of likelihood is obtained. Akaike Information Criterion (AIC) was used to judge the performance of the model. Smaller the AIC value, the better the model. AIC=-2LogL+2K ------------------------------(3) Where Log L is the log likelihood; K is the number of estimated parameters. VIII. VARIABLES SELECTION Various combinations of the variables selected were tried the best combinations are tabulated in table 2. Accident rate per year per Km was chosen as dependent variable. Crash rate is defined as CR i = TA i / L i / NY. Where CR i =Crash Rate on segment i TA i = Total accidents on segment i L i = Length in Km. of segment i NY= Number of Years Fig.5 Shoulder Width vs Crash Rate 70

Table.2 variables used in models Model No. Independent Variables Dependent Variable 1 Access Density(AD), Shoulder Width(SW) 2 Access Density(AD), Shoulder Width(SW), Annual Average Daily Traffic(AADT) 3 Shoulder Width(SW), Annual Average Daily Traffic(AADT), Lane Width(LW) 4 Access Density(AD), Shoulder Width(SW), Lane Width(LW) Accidents/Yr/Km (Crash Rate) 5 Access Density(AD), Shoulder Width(SW), Annual Average Daily Traffic(AADT), Lane Width(LW) IX. RESULT Result is given in Table.3 The final model is selected on the basis of Akaike Information Criteria A. Final Model ACC/YR/KM = 0.167+0.233*AD-0.118*SW -0.116*LW+0.0001685*AADT Predicted Frequency = 1 1 y y ε* i 1 i i 1 1 1 y 1 i i i Where is Exposure term = Total length of Study area in Kilometers α is overdispersion Parameter =0.35 (for present data collected ) B. Model Testing: The model finally selected is tested for the accident data collected for the year 2010. The testing result is given in table.4 X. CONCLUSIONS i. Pedestrian safety is greatly influenced by number of access points per unit length of the road. Each additional access point per kilometer of road length may increase accident rate by more than 100%. ii. Shoulder width is also a highly influential factor affecting pedestrian safety. Additional 1m width of shoulder will reduce the accident by 50%. iii. Additional lane width of 1m may reduce the accident by 50%. iv. Predicted frequency has very good correlation with the observed frequency XI. REMEDIES i. Access to main highways should be properly designed. ii. Sufficient shoulder width or provision of safe walking places along the highways should be provided. iii. Elevated and visible designated areas for crossing of roads in all possible places. iv. Separation of pedestrian movement from heavy moving traffic in all possible places. v. Speed control by road design, traffic calming and enforcement on highways, near traffic generators like educational institutions, business places and hospitals. 71

Table.3 Parameter estimates for different variables used in model (Using SPSS) Variable Model 1 Model 2 Model 3 Model 4 Model 5 Intercept 0.366-0.769 1.378 1.866 0.167 AD 0.329 0.238 0.315 0.233 SW -0.139-0.121-0.270-0.132-0.118 AADT 0.000172 0.000282 0.0001685 LW -0.264-0.191-0.116 AIC 29.61 5.584 47.08 29.506 4.206 Table.4 Predicted and Observed frequency Accident Frequency Probability Predicted Frequency Observed Frequency 0 0.355 35.5 46 1 0.308 30.8 16 2 0.18 18 10 3 0.0888 8.88 4 4 0.0395 3.95 2 5 0.01645 1.645 3 6 0.00654 0.654 1 7 0.0025 0.25 0 8 0.0009 0.09 0 9 0.00034 0.034 0 10 0.000124 0.0124 0 R-Square 0.795 vi. Pedestrian education programmes for safe walking, vii. Implementing pedestrian safety programmes require a skillful mix of road engineering and enforcement measures along with education for people to accept changes. XII. REFERENCES [1]. Road accident in India(Ministry of Road Transport and Highway)2010 [2]. Zegeer, C V., Deen, R. C. & Mayes, J. G. Effect of lane and shoulder widths on accident reduction on rural two-lane roads. TRR 806 1981.pp 33-43 [3]. Masaeid, H.R. & Nelson, C.D., Pedestrian accidents and their relationship to street geometrics and operation variables in Jordan, Indian Highways, Vol. 24, No. 9, Sept. [4]. Karlaftis, M.G. & Golias, I., Effect of road geometry and traffic volumes on rural roadway accident rates, Accident Analysis and Prevention vol.34, Issue 3, 2002, pp. 357-365. [5]. Okamoto, A Method to cope up with random errors of observed accident rates Safety literature, Vol.4, page 317-332, 1989 [5]. Garber, N.J., Wu, L, Stochastic models relating crash probabilities with geometric and corresponding traffic characteristics data, Research report No UVACTS-5-15-74 Center for transportation studies at the University of Virginia 1989 [6]. Poach, M. & Mannering, F., Negative binomial analysis of intersection accident frequencies ASCE s Journal of Transportation Engineering Vol.122 No. 2 March April 1996, pp. 105-113. [7]. Miaou, S.P. The relationship between truck accidents and geometric design of road sections: Poisson versus negative binomial regressions, Accident Analysis and Prevention vol. 26, issue 4, 1994, pp. 471-482 [8]. Landge, V.S., Jain S.S. & Parida, M., Modeling traffic accidents on two lane rural highways under mixed traffic conditions, 87th Annual Meeting of Transportation Research Board, Jan 2006. (CD Rom) 72

Biography: First Author:A.K.Sharma,born on 28-12-64,graduated in civil Engineering from NIT Raipur(Chattisgarh,India) in 1987. He completed his post-graduate in Highway and Traffic Engineering from IIT Kharagpur (India) in 1989. Presently working at Shri Ramdeobaba College of Engineering and Management, Nagpur, as Associate professor in Civil Engineering Department. His main field of work has been pavement and traffic engineering. He has guided many undergradute and post graduate dissertations. Presently he is persuing his Doctoral research in the field of traffic safety. Second Author: V.S.Landge, born on 02-10-68, graduated from RKNEC, Nagpur (India) in 1991. He completed his post graduate studies in Transportation Engineering from BITS Pillani(Rajasthan, India) in 1993. He completed his Doctoral degree form IIT Roorkee(India) in 2006. Presently working at VNIT, Nagpur as Associate professor in Civil Engineering Department. His main field of work has been traffic engineering particularly in the field of traffic Safety. He is a member of the State Technical Agency for Pradhan Mantri Gram Sadak Yojna for Vidarbha Region in India He has guided many undergraduate and post graduate dissertations. 73