Improving Road Safety by Profiling Different Accident Type Te Team - PowerPoint PPT Presentation

Improving Road Safety by Profiling Different Accident Type Te Team am 7 : 7 : An Ange gela, la, Ayl ylada, ada, Ce Celi lia, a, Dobby obby, , Ma Mahs hsa

DATA MINING GOAL Profile certain accident types by identifying the factors that might be the cause of that accident. Accident types we are focusing on: FrontHit, SideHit, BackHit, PedCrossing, and Scratch 2

BUSINESS GOAL To assist Transportation Department of Taipei City Government(Client) to improve the road safety in Taipei with more efficient way by learning the conditions that distinguished different type of accidents. To help the government decrease in number of certain accident types while wisely use the budgets. To have a safer road usage for Taipei citizens. 3

DATA DESCRIPTION Data set: Accident records in Taipei from Data Taipei Time variables Environmental variables 4

DATA DESCRIPTION (Cont.) 30 columns → 17,991 2013 2013 - 39,577 39,577 rows 30 17,991 rows 11 11 columns 30 columns → 17,843 2012 2012 - 39,062 39,062 rows 30 17,843 rows 11 11 columns 30 columns → 18,963 2011 - 41,082 2011 41,082 rows 30 18,963 rows 11 11 columns We combined the records of same case ID one row: one record => one row: one accident case We also deleted records that has accident on highway. (not our goal) To reduce dimensionalities, we have - deleted some columns that are not significant in profiling accident types - binned the categories that are similar 5

DATA DESCRIPTION (Cont.) We found that our data is imbalanced To deal with this, we need to oversample our data. 6

METHODS Disc Di scri rimi minant Anal nant Analys ysis is (R (R) Decis ision ion Tr Tree e (R (R) Decis ision ion Tr Tree e (X (XLm Lmin iner) r) Ran andom dom Fo Fores rest t (R (R) Decis ision ion Tr Tree e (R (Rapi apidMi dMiner) ner) 7

Method Chosen- Decision Tree (RapidMiner) Main process Data Preparation (and Oversampling) Sub-process 8

Method Chosen- Decision Tree (RapidMiner) Under most of the circumstances, all kinds of accidents could happen.

Results However, there are circumstances where only certain accident type happens. Out utco come me Fa Factors ctors Fa Factors ctors 1. 1. At At Cir ircl cle/ e/Pl Plaza, aza, Spee eed d li limi mit <= <= 65 65, , 1 Acci Ac cident dent Fr Front ontHit Hit Fa Fast/ st/Slow low/Normal Normal La Lane ne Lo Location cation 1. Su 1. Sunny nny, , Sp Speed eed li limi mit > 65 > 65, , No Non-For Fork, k, No Non-Cross ross Road ad 2 Spee eed d Li Limi mit 2. Other Road Type, Speed limit <65, Non-Traffic Signal Ba Back ckHit Hit 3. Other Road Type, Speed limit <65, Motorcycle Lane 3 Road ad Ty Type 4. At Circle/Plaza, Motorcycle Lane 4 Sig igna nal 1. 1. Spee eed d li limi mit 65 65~7 ~75, 5, wh when en p pav avement ement is is We Wet, t, Sid ideH eHit it at Cross at ss Road ad 5 Weathe We ather 1. At Zebra Crossing 6 We Weekday/ ekday/ Pe PedC dCrossin ossing 2. 2. At At Cir ircl cle/ e/Pl Plaza, aza, Sid idew ewal alk, k, Ni Nigh ght Ti Time me, , Weekend We ekend Spee eed d li limi mit <= <= 75 75 1. Speed limit >85 Scr cratc atch 2. Speed limit <=85, Rainy, Weekend 10

Performance Evaluation 1. Cr 1. Cross Vali ss Validatio dation n (w (wit ith 10 Va h 10 Vali lidati dations ons) Those cannot be distinguished very well are predicted as SideHit Ov Overal erall Ac l Accurac uracy: y: 22 22.7 .79% 9% Clas Cl ass o s of f In Inte tere rest st Accurac uracy: y: 27 27.3 .34% 4% Therefore lower 1. Pur 1. Purit ity o y of f the the end end no nodes des the accuracy We did identify some accident types that happens under certain circumstances 11

RECOMMENDATION 1. Build a br brid idge ge or und under ergr groun ound d wa walk lkwa way for pedestrian at the circles/ plaza to increase their safety. 2. Install more signals at the circles/ plaza to avoid front hit. 3. Install anti-slip textures and decrease the speed limit before the cross roads to decrease the number of side hit and scratch. 4. Dynamic speed limit to avoid backhit. (low speed limit in sunny days ?) 12

13