Application of data mining in road traffic accident analysis

Abstract

The existing road traffic accident (RTA) analysis system in Bangladesh is more focused onto record management and basic data analysis i.e. characteristics analysis purposes rather than using it as a source of intelligence. Although MAAP based accident database constitute the respiratory for RTA information of the country, its application is constrained by a number of limitations. However, most of the previous studies focused on a few risk factors, some specific road users or certain types of crashes; and therefore the important factors affecting injury or crash severity have not been completely recognized yet. Data mining (DM) has the potential to eliminate RTA data related deficiencies as well as statistical limitations. Even DM is able to quantify multiple relationships, which provides the insight for policy level decisions. Therefore, DM has been utilized in this thesis to elicit reasonable, novel, and interesting facts and also to confirm some perceived facts using RTA data (1998-2010) from ARI, BUET. Several DM algorithms have been adopted for the study. At first, hierarchical clustering (HC) methodology was employed to form natural data groups and to identify hazardous clusters; then random forest (RF) was applied to identify, rank, and thus select a subset of variables from a large variable space, to be considered for this study. Finally, classification and regression trees (CART) have been allowed to investigate the accident severity mechanism of the hazardous clusters. Nearly 10 percent of the pedestrian accidents are triggered by other accident/collision types, which indicate that may be pedestrians are not only the victims but also a stimulating factor for some accidents. Dividers in urban areas have been found quite effective in reducing fatal (38.23% fatal vs 57.78% fatal where there are no dividers) pedestrian accidents. Traffic control systems especially police controlled traffic control system in urban areas have been identified as persuasive in reducing pedestrian fatal accidents (in some cases 0% fatal incidences). Geometric sections without police controlled traffic control system have been acknowledged as a bracing factor for fatal pedestrian accidents. Straight and flat geometric sections of roadways have generated more double vehicle fatal accidents (more than 80% accidents are fatal) than other types (e.g. curve only, slope only, curve and slope and crest) of geometric sections (nearly 70% fatal accidents). The latter part of the previous finding got worse when the sections were associated with head on, right angle, overturn, hit object in road and hit animal type of collisions (76.22% fatal); or occurred on national and regional highways or feeder roads (71% fatal); or during dawn/dusk and night (unlit) lighting condition (90.91% fatal); or in daylight or night (lit) light condition but with no or centerline marking traffic control system (75.21% fatal). Head on, right angle, side swipe, hit object in road, and hit object off road collision types affiliated with curve only, slope only, and curve and slope geometric sections of the roadways produced 85.29 percent fatal single vehicle crashes. Dawn/dusk and night (unlit) lighting condition attributed 87.88 percent single vehicle fatal accidents. Brick and earthen road surfaces have generated 86.67% fatal single vehicle crashes even in daylight and night (lit) condition. On the contrary, sealed surface even affiliated with rainy weather has ensued less fatal single vehicle crashes (58.82% non-fatal crashes).Wet and flooded surface conditions of roads have resulted in 94.74 percent fatal single vehicle crashes. Nevertheless, one-way routes concomitant with dry and muddy surface prompted only 20 percent fatal cases as always perceived; whereas in case of two-way roads it shoots up to 86.54 percent fatal single vehicle accidents.