Accident mapping algorithm results

AMF was implemented to all the 10,520 STATS19 accidents. It has been found that the time required to process 10,520 accidents is 230 minutes (by using a laptop PC with 4GB RAM and 3.4GHz processing speed). This suggests that the developed method can process 46 accidents per minute. The accuracy of the four algorithms (AMM1, AMM2, AMM3, AMF) was then evaluated using the reference 716 accidents discussed in section 3.3.5; the accuracy that was estimated for each algorithm was the percentage of refer- ence cases that were assigned to the same road segment with the one that was selected manually. The percentage of accuracy was estimated at the 95% confidence level with a confidence interval of± 1.1%. All the results can be found at Table 4.5. The total accuracy levels were found to be 81.6%, 87.7%, 85.0% and 98.9% for the AMM1, AMM2,

reported road types (roundabouts, slip roads and main carriageways respectively) were also estimated in order to identify possible weaknesses of the method in the identifica- tion of specific categories of road accidents. It was revealed that AMM1 and AMM3 face particular difficulty in identifying the correct location for accidents that occurred on roundabouts and that AMM2 has the same problem for main carriageway accidents. In contrast, AMF gives accurate results for roundabout and slip road accidents whereas, for main carriageway accidents the results are less accurate. The mismatches on main car- riageways are mainly due to the errors of the reported road name and type of the accident database. The mean Distance for each of the methods was calculated and found to be: 4.43 m for AMM1, 6.47 m for AMM2, 14.54m for AMM3 and 8.53 m for AMF.

It is clear that the AMF method with error just above 1% gives the most reliable matching results among the examined methods. An interesting outcome is that the correct segment is not always the closest to the reported accident location that is in line with other studies (Loo, 2006; Deka and Quddus, 2014; Imprialou et al., 2015). An additional result to that was that even the closest segment that has the same road name and road type with the examined accident can be erroneous. This highlights the importance of the intended ve- hicle direction as a variable for an accident mapping algorithm. However, from the results of the AMM3 method that considers the vehicle’s intended direction it is revealed that the inclusion of this variable in an inflexible formula does not guarantee the accuracy of the results. This supports the selection of fuzzy inference systems for accident location identification that provide a flexible framework that adapts to the reported data of each case individually. For the AMF method, the 99th percentile of the Distance was found to be 56.8 m and the 98.5th percentile 49.9 m confirming the validity of the selection of the 50 m threshold boundary. In other words, the fact that 98.5% of the cases have Distance less than 49.9 m justifies the need for a manual check in order to confirm the accuracy of the segment selection (as it is described at the Additional Steps 1 and 3 in section 3.3.4)) when the Distance from the selected segment is over 50 m.

Table 4.5: Estimation of accuracy and average distance for the four examined accident mapping methods

Total Accuracy Average Method Roundabouts Slip Roads Carriageways (%) Distance

(%) (%) (%) (95%,1.1) (m)

AMM1 74.0 80.3 82.3 81.6 4.63

AMM2 98.0 96.7 86.0 87.7 6.47

AMM3 52.0 96.7 88.3 85.0 14.54

AMF 100.0 100.0 98.7 98.9 8.53

From the 10,520 cases that were matched with the AMF method, there were 266 (2.5%) that needed manual checking according to the additional steps 1 and 2. 14 of them were accidents that could not be matched with any segment of the database due to simultane- ous road name and road type mismatch with all the potential candidate segments. From the 266 cases that were checked manually, 107 (1%) needed manual correction. From the entire database, there were overall 206 (2%) cases of accidents that were matched to road segments that had different road names than the reported, and 557 (5.3%) segments that had different road types than reported. After the 107 manual corrections, there were 36 (0.3%) cases where neither the road name nor type of the selected road segment was the same with those referred to on the accident report. This situation indicates the existence of some inconsistency between the network and the accident database that is mostly re- sponsible for the estimated error of the developed method.

Figures 4.11 and 4.12 represent graphically the accident locations when they are superim- posed on the digital road network; before accident-mapping (a) and after the implemen- tation of the AMM1(b), AMM2 (c), AMM3(d) and AMF(e) algorithms respectively. It can be easily noticed that the majority of the reported accident locations (Figures 4.11(a) and 4.12(a)) do not fall exactly onto a road segment and some of them are placed between two or more road sections. The accident locations indicated by the four methods (Figures 4.11(b)-4.11(e) and 4.12(b)-4.12(e)) have both similarities and differences. The locations of AMM1 and AMM2 (Figures 4.11(b), 4.11(c), 4.12(b) and 4.12(c)) are very similar to each other, as it was expected, but they are quite different from the AMF (Figures 4.11(e)

and 4.12(e)) locations as the accidents are not necessarily assigned to the closest road seg- ment, but the one that has the highest Matching Score. The locations of the AMM3 are almost identical to those indicated by AMF on the main carriageway accidents (Figure 4.11(d)) however the locations for the roundabout accidents (Figures 4.12(d)) are very different as AMM3 was found to have only 52.0% of accuracy for roundabouts.

(a) STATS 19 reports (unmatched)

(b) AMM1 (c) AMM2

(d) AMM3 (e) AMF

(a) STATS 19 reports (unmatched)

(b) AMM1 (c) AMM2

(d) AMM3 (e) AMF