Data Mining of Driver Characteristics to Spatial and Temporal Hotspots of Single Vehicle Crashes in Western Australia

Transcription

1 19th Internatinal Cngress n Mdelling and Simulatin, Perth, Australia, December Data Mining f Driver Characteristics t Spatial and Tempral Htspts f Single Vehicle Crashes in Western Australia Jianhng (Cecilia) Xia Department f Spatial Sciences, Curtin University, Australian c.xia@curtin.edu.au Abstract: This paper presents innvative methds fr identifying the characteristics f drivers invlved in single vehicle crashes (SVCs) in Western Australia when viewed spatially and temprally. Spatial and tempral htspts f SVCs can be defined as clusters f SVCs that have exceeded the expected value ver a certain time perid and at certain lcatins. The EM (Expectatin-Maximisatin) algrithm was adpted t identify the characteristics f driver invlved in vehicle crashes. Drivers were divided int different segments based n their sci-demgraphic characteristics, i.e., age and gender and driver related crash factrs such as, drive license s types, alchl cnsumptin, speeding, fatigue and inattentin etc. The spatial htspts f the SVCs were identified using the Kernel Density Estimatin methd. Cmap was used t integrate space, time and characteristics f drivers int ne view t better understand the nature f the SVCs. The main cnclusin draws frm this study are as fllws: Characteristics f drivers subject t SVCs Change ver time with different cntributing factrs. Sme subgrups f drivers at greater risk were identified. Females aged arund 19, with BACs arund 0.03%, n prbatinary licences were mre likely t be invlved in crashes between the hurs f midnight and 2:10am and require hspital attentin. Males aged arund 61, n full licences were mre likely t be invlved in crashes between the hurs f 7:55am and 10:55am because f inattentin and require medical r hspital treatment. Males, and sme females, aged arund 26, n full licences, and with BACs arund 0.07% were mre likely t be invlved in crashes between the hurs f 2:10am and 4:55am because f alchl, inattentin, and speeding, and require hspital treatment. Males, and sme females, aged arund 36 were mre likely t be invlved in crashes between the hurs f 10:55am and 17:40am because f inattentin and require medical attentin. Males and females aged arund 17 n prbatinary licences were mre likely t be invlved in crashes between the hurs f 17:40pm and 20:55pm because f inattentin, and require medical attentin. Alchl and speeding were fund t be the main cntributing factrs in SVCs at night r early mrning, and inattentin was the main factr in SVCs during the day. Severity f SVCs is generally higher at night than during the day. The distributin f htspts f SVCs varies ver different time perids The results f this research will help rad safety management authrities better understand the underlying nature f why crashes happen and where and when they d s that remedial actin can take int accunt the characteristics f drivers. Keywrds: Vehicle crash, spatial and tempral htspt, cluster and utlier analysis, driver characteristics, expectatin-maximisatin (EM) algrithm 697

2 Xia., Data Mining f Driver Characteristics t Spatial and Tempral Htspts f Single Vehicle Crashes 1. INTRODUCTION The rate and severity f vehicle crashes have cnsistently been fund t be assciated with certain driver characteristics. Fr example, Bédard et al. (2002) fund that male drivers yunger than 30 years dminate the fatality recrds. Hwever, lder drivers are mre likely t be fatally injured as a result f the actins f ther drivers than are yunger drivers (Bufus et al., 2008). Vaez and Laflamme (2005) extended age and gender t ther demgraphic attributes such as class f rigin and educatinal attainment and suggest sme at-risk ppulatins are males, persns aged 18 19, thse frm husehlds classified as wrkers and thers (including, e.g. the lng-term unemplyed and thse n lng-term sick leave), and thse with lw educatinal attainment. Incme and race were als studied by Lerner et al.(2001) t understand their influence n seatbelt use by adults injured in mtr vehicle crashes. The results shwed that lder persns, wmen, Caucasians, and individuals with greater incmes tended t wear seat belts mre. Hwever, Dugherty, Pless and Wilkins (1990) fund that Mntreal neighburhds with the lwest quintile f incme levels had fur-times as many fatal vehicle crashes invlving children than ther mre advantaged neighburhds. Cliftn et al. (2009) cnfirmed these previus demgraphic studies, but als discvered that lighting cnditins, the time f the day, transit access, and pedestrian cnnectivity are assciated with injury severity as a result f crashes. This research tends t suggest that lcatin matters in vehicle crashes. Further, Plug et al. (2011) applied different spatial and tempral visualisatin techniques t understand the prcess underlying single vehicle crash patterns and their causes and shwed that time as well as lcatin influence vehicle crashes. Accrding t Tbler s first law f Gegraphy: Everything is related t everything else, but near things are mre related than thse far apart (Tbler, 1970). When applied t vehicle crashes, thse in clse prximity t each ther prbably have similar causes. Fr example, Hall and Zadr (1981) fund clusters f crashes htspts ccurred n certain types f rads because f the gemetry f the rads, such as severe hrizntal curves (average 1.7 degree at crash sites). Similarly, at certain htspts, the driver characteristics may be a factr. Based n this thery, this paper aims t identify whether the characteristics f drivers invlved in vehicle crashes change acrss different lcatins and ver time. Cmap will be used t identify spatial and tempral patterns f vehicle crashes, and the EM (Expectatin-Maximisatin) algrithm will be used t identify characteristics f drivers assciated Figure 1. The study area with these spatial and tempral patterns f vehicle crashes. 2. METHODS 2.1. Data cllectin and study area The data used in this paper were btained frm a database maintained by the Main Rads Department in Western Australia and which cntains infrmatin abut all plice-reprted vehicle crashes in Western Australia. Accrding t Plug et al. (2011), the Perth Lcal Gvernment Area (LGA) is a better scale t investigate htspts spatially. Therefre, we cncentrated n the characteristics f drivers invlved in single 698

3 Xia., Data Mining f Driver Characteristics t Spatial and Tempral Htspts f Single Vehicle Crashes vehicle crashes (SVCs) between 1999 and 2008 in this study area. The study area and the distributin f all crashes are illustrated in Figure 1. Table 1 summarises the percentage frequency f single vehicle crashes in the Perth LGA frm 1999 t The nature values 7 t 10 represent SVCs Characteristics f drivers Table 1. Summary f SVCs and MVCs in Perth LGA Value Crashes by Nature Frequency % f Ttal Hit Pedestrian Hit Animal Hit Object Nn Cllisin The sci-demgraphic variables included in the analyses are shwn in Table 2. Table 2: Sci-demgraphic variables. Variable Range and Ntes. Sex Age Prtectin Driver s Licence Type Injury type Male r female The age range in the data set was frm 9 t 94 years ld. Seat belt wrn r nt Licenced, suspended, learner, cancelled, unlicenced, prbatinary Death within 30 days f accident; injured and requiring medical attentin but nt in hspital; injured and admitted t hspital; and n medical attentin required (Main Rads Western Australia, 2011). Bld Alchl Cntent The bld-alchl-cntent (BAC) range in the data set was frm 0.0 t Under Western Australian law, drivers with a full licence may be issued an infringement ntice if their BAC exceeds 0.05%; fr drivers with a prbatinary licence, it is 0.02% (Western Australia Plice, 2011). Accident Cause (as Speeding, fatigue, inattentin, alchl assessed by the attending plice fficers) 2.3. Kernel Density Estimatin The Kernel Density Estimatin technique (KDE) is a well-established methd fr identifying htspts frm pint data (see O'Sullivan and Unwin, 2003 fr detailed examples). It creates a field representatin f htspts using different clur schemes. The density f pints fr each pixel in the utput raster is estimated by cunting the number f pints in a defined regin r kernel, divided by the area f the kernel. Thus, a discrete density surface is made cntinuus by interplatin and, fr example, darker clurs highlight lcatins f high intensity. Fr this study, KDE was used t estimate the density f SVCs (r htspts) in the Perth LGA. The Spatial Analyst KDE tl in ESRI s ArcGIS 10 sftware was used t calculate the KDE (ESRI, 2010) Cmap Cmap, a visualisatin tl, was develped frm a cncept f cplt (Tufte, 1990), which is based n principle f dividing a whle int subsets, and then pltting the subset data t investigate differences between individual subsets. Cmap has been used t investigate spatial and tempral patterns f phenmenn such as disease, crime r vehicle crashes (Asgary et al., 2010; Plug et al., 2011). A Cmap is created this way: 1. Run the classificatin f time pint data f the vehicle crashes using a natural break scheme (Jenks, 1967). This methd is designed t arrange time pints int a grup with minimum deviatin frm the grup mean, while seeking t maximise the deviatin frm mean f the ther grups. Therefre the time pints in the same grup shuld have a similar nature. 699

4 Xia., Data Mining f Driver Characteristics t Spatial and Tempral Htspts f Single Vehicle Crashes 2. Break the vehicle crash data int eight subsets based n the classified time intervals because f the structure f the Cmap. 3. Use the KDE functin t identify spatial patterns f vehicle crashes. 4. Create maps f the spatial pattern f vehicle crashes fr each subset. 5. Illustrate this sequence f maps with defined time intervals using the Cmap structure. In this study, we applied Cmap t identify changes in the SVC spatial patterns ver different time intervals EM algrithm (Expectatin and Maximisatin) The EM algrithm (Dempster et al., 1977) is a mdel-based clustering methd t build certain mdels fr clusters based n parametric distributins, like a Gaussian (cntinuus) r a Pissn (discrete). Each individual cluster can be represented by a distributin. The whle data set is a mixture f these individual distributins. The gal f the algrithm is t ptimise the fit between the mdels and data, which means that each distributin can capture dminant patterns f data. An advantage f the EM algrithm is that it can be used fr bth numeric (cntinuus) and categrical (discrete) data. Detailed infrmatin regarded t EM algrithm can be fund frm Xia et al (2010). This paper used the EM algrithm t identify characteristics f divers based n their scial-demgraphic variables defined in sectin 2.2. Age and BAC are cntinuus variables and ther variables are discrete. Drivers were gruped int the same class if they have similar characteristics. The number f grups f drivers will be decided by the nature f data. Mre than ne grup f drivers culd be identified fr a time interval. 3. RESULTS Figure 2 illustrates the spatial and tempral pattern f single vehicle crashes by Cmap. The time pints f the SVCs were classified int eight time intervals using Jenks Natural break classificatin. Figure 3 cnfirms that the classificatin results are crrect. Each time interval cvers ne r tw spaced peaks. In Figure 2, each bar at the tp represents the length f a time interval marked by the end f a time instant. Fr example, the first time interval is between midnight and 2:10. Each map illustrates the spatial Figure 2. The Cmap f the SVCs. in the Perth LGA frm 1999 t 2008 distributin f the SVCs in Perth LGA within the classified time interval given by the same numbering system as time bar. Fr example, Map 1 shws the distributin f SVCs between midnight and 2:10am. Generally, htspts f SVCs ccur mre frequently arund intersectins and alng sme majr streets. Hwever, they vary at different lcatins during different time intervals, which are prbably due t variatins f pedestrians activities. 700

5 Xia., Data Mining f Driver Characteristics t Spatial and Tempral Htspts f Single Vehicle Crashes The characteristics f drivers als change ver these time intervals. Fr example, males are generally mre likely t be invlved in SVCs than females. Hwever, mre females aged arund 19 tend t be severely injured at night where the crash is attributed t alchl (BAC arund 0.03%), speeding, r inattentin. Males aged arund 30 are mre likely t speed between midnight and 2:10am and be invlved in crashes that require hspital attentin (see Figure 4). Between 2:10am and 4:59am, mst f thse invlved in SVCs were aged arund 26 and alchl, speeding, and inattentin are the main causes. The injuries tend t be mre serius and mst drivers have t attend hspital. Males are fur times mre likely t be invlved in SVCs than females during this time interval, but females aged arund 22 and with BACs arund 0.07% tend t be seriusly injured. Figure 3. The frequency f the SVCs ver time During the early mrning interval, between 5:00am and 7:55am, the severity f injuries decreases, but speeding and alchl are still the main causes. Hwever, the reprted BACs were lwer than fr the immediately prir time interval, meanwhile, age and gender types were similar fr the tw perids. Between 7:55am and 10:55am, mst SVCs are caused by inattentin and resulted in injuries requiring medical attentin. Three vulnerable grups were identified: yung driers aged arund 17, male drivers aged arund 61, and drivers aged arund 35. Figure 4. The Characteristics f drivers 701

6 Xia., Data Mining f Driver Characteristics t Spatial and Tempral Htspts f Single Vehicle Crashes Between 10:55am and 14:15pm and 14:15pm and 17:40pm, drivers invlved in SVCs are mainly aged arund 36 r 37, required medical attentin, and the cause was attributed t inattentin. As Evening wre n, yunger and lder drivers became invlved in SVCs and alchl and speeding re-emerged as the cause. 4. DISCUSSIONS AND CONCLUSIONS The EM algrithm is widely used in clustering analysis (Laird, 2010; Xia et al., 2010) and has the advantage f being suited t bth numeric and categrical data. S it was adpted fr this study. Hwever, there are tw issues that arise when the EM algrithm is applied in practice: Because the initialisatin is randm, the algrithm may cnverge t lcal likelihd maxima instead f the glbal maximum. This can be avided by running the algrithm many times, and chsing the mdel results that crrespnd t the glbal maximum. The EM algrithm cannt be used t cmpute the number f clusters because the likelihd f a mixture mdel always increases as the number f clusters (and number f parameters) increases. T select the ptimal number f clusters, the likelihd must be penalised as the number f mdel parameters. Cmap has been used in this study t illustrate the spatial and tempral distributin f SVCs in the Perth LGA and the variatin f characteristics f drivers subject t SVCs ver time. Further research will lk at the characteristics f drivers at sme majr htspts instead f the whle regin t get a better idea f the true spatial, tempral, and characteristics f drivers 3D interactins. The time clustering methd used in this study is Jenks natural break classificatin scheme. It prved t be an effective classificatin methd t break dwn time pint int different intervals based n their frequency distributin. We limited the class number t eight because f Cmap restrictins, but in practice, the number f classes shuld be determined by the nature f the data. Therefre, in the future, we will investigate better classificatin methds and mdify the Cmap methd t make it mre flexible. The main cnclusin draws frm this study are as fllws: The characteristics f drivers invlved in SVCs change ver space and time with different cntributing factrs. Sme subgrups f drivers at greater risk were identified: Females aged arund 19, with BACs arund 0.03%, n prbatinary licences were mre likely t be invlved in crashes between the hurs f midnight and 2:10am and require hspital attentin. Males aged arund 61, n full licences were mre likely t be invlved in crashes between the hurs f 7:55am and 10:55am because f inattentin and require medical r hspital treatment. Males, and sme females, aged arund 26, n full licences, and with BACs arund 0.07% were mre likely t be invlved in crashes between the hurs f 2:10am and 4:55am because f alchl, inattentin, and speeding, and require hspital treatment. Males, and sme females, aged arund 36 were mre likely t be invlved in crashes between the hurs f 10:55am and 17:40am because f inattentin and require medical attentin. Males and females aged arund 17 n prbatinary licences were mre likely t be invlved in crashes between the hurs f 17:40pm and 20:55pm because f inattentin, and require medical attentin. Alchl and speeding were fund t be the main factrs assciated with SVCs at night r early mrning, and inattentin was the main factr in SVCs during the day. The severity f SVCs was generally higher at night than during the day. The distributin f SVC htspts mved ver time. In the future, we will cnduct similar analysis fr different types f SVCs, such as hit-pedestrian crashes r multi-vehicle crashes. It will be interesting t identify the characteristics f drivers in relatin t vehicle 702

7 Xia., Data Mining f Driver Characteristics t Spatial and Tempral Htspts f Single Vehicle Crashes crashes at majr htspts and especially include sme envirnmental factrs, such as weather cnditins, in the analysis. Fr this study, we chse the Perth LGA, but further research will be undertaken in ther areas, such as turist twns r residential and industrial areas, t see if the same results hld. ACKNOWLEDGMENTS We wish t acknwledge the assistance f Main Rads WA in prviding access t the WA rad crash data files needed fr this study. Thanks are extended t Craig Caulfield fr editing the paper, Matthew Legge fr discussing sme issues related t the paper and tw reviewers fr their cmments that help imprve the manuscript. REFERENCES Asgary, A., Ghaffari, A., & Levy, J. (2010). Spatial and tempral analyses f structural fire incidents and their causes: A case f Trnt, Canada. Fire Safety Jurnal, 45(1), Bédard, M., Guyatt, G. H., Stnes, M. J., & Hirdes, J. P. (2002). The independent cntributin f driver, crash, and vehicle characteristics t driver fatalities. Accident Analysis & Preventin, 34(6), Bufus, S., Finch, C., Hayen, A., & Williamsn, A. (2008). The impact f envirnmental, vehicle and driver characteristics n injury severity in lder drivers hspitalized as a result f a traffic crash. Jurnal f Safety Research, 39(1), Cliftn, K. J., Burnier, C. V., & Akar, G. (2009). Severity f injury resulting frm pedestrian-vehicle crashes: What can we learn frm examining the built envirnment? Transprtatin Research Part D: Transprt and Envirnment, 14(6), Dempster, A. P., Laird, N. M., & Rubin, D. B. (1977). Maximum likelihd frm incmplete data via the EM algrithm.. Jurnal f the Ryal Statistical Sciety, Series B, 34, Dugherty, G., Pless, I. B., & Wilkins, R. (1990). Scial class and the ccurrence f traffic injuries and deaths in urban children. Canadian Jurnal f Public Health, 81(3), ESRI. (2010). ArcGIS 10. New Yrk: ESRI. Jenks, G. F. (1967). The Data Mdel Cncept in Statistical Mapping. Internatinal Yearbk f Cartgraphy(7), Hall, J. W., & Zadr, P. (1981). A survey f single vehicle fatal rllver crash sites in New Mexic. Transprtatin Research Recrd (819), 1-8. Laird, N. M. (2010). The EM Algrithm in Genetics, Genmics and Public Health. Statistical science, 25(4), Lerner, E. B., Jehle, D. V. K., Billittier, A. J., Mscati, R. M., Cnnery, C. M., & Stiller, G. (2001). The influence f demgraphic factrs n seatbelt use by adults injured in mtr vehicle crashes. Accident Analysis & Preventin, 33(5), Main Rads Western Australia. (2011). Vehicle crash dictinary. O'Sullivan, D., & Unwin, D. J. (2003). Gegraphic Infrmatin Analysis: Jhn Wiley & Sns, Inc. Plug, C., Xia, J., & Caulfield, C. (2011). Spatial and tempral visualisatin techniques fr crash analysis. Accident Analysis & Preventin, In Press, Crrected Prf. Tbler, W. R. (1970). A Cmputer Mvie Simulating Urban Grwth in the Detrit Regin. Ecnmic Gegraphy, 46, Tufte, E. R. (1990). Envisining infrmatin. Cheshire, Cnnecticut: Graphics Press. Vaez, M., & Laflamme, L. (2005). Impaired driving and mtr vehicle crashes amng Swedish yuth: An investigatin int drivers' scidemgraphic characteristics. Accident Analysis & Preventin, 37(4), Western Australia Plice. (2011). Drink driving penalities. Retrieved 2 July, 2011, frm Xia, J., Evans, F. H., Spilsbury, K., Ciesielski, V., Arrwsmith, C., & Wright, G. (2010). Market segments based n the dminant mvement patterns f turists. Turism Management, 31(4),