Revisiting Traffic theories for different models in lights of Pre-Crash Traffic Condition

Abstract

The concept of predicting the probability of a road crash on an access controlled road is gaining momentum in recent years. Considerable research has been carried out to establish several statistical and artificial intelligence based models which can take real-time traffic data as input to predict the short time crash probability. In general, the underlying methodology builds knowledge about traffic condition during the time periods when there is no crash taking place from the historical detector data and based on that, if they find the real-time traffic data to be differing significantly from the normal data, they classify it as crash prone. Interestingly, although these models heavily deal with the basic traffic flow variables, such as, speed, flow, density, and/or their descriptive statistics or surrogate measures to build and operate such models, very few studies have been focused on whether and how the fundamental relationships of traffic flow theories differ for pre-crash and normal traffic conditions. The main purpose of this research is to revisit classical two-phase traffic flow theories in lights of pre-crash and normal traffic conditions and identify the variations in traffic conditions using pre-crash and normal traffic data. The data for the analysis was extracted from 210 detectors from March 2014 to August 2014 on the Shibuya 3 and Shinjuku 4 routes of Tokyo Metropolitan Expressway in Japan. The detector data consisted of information on speed, vehicle count, occupancy and number of heavy vehicles aggregated for each minute for each lane of the study area. During this time 620 crashes took place. Matching the crash and the detector data, pre-crash and normal traffic condition datasets were prepared. For pre-crash data, for each crash point, data for each minute for the ten minutes leading to crash were collected from the nearest detector, nearest upstream and nearest downstream detectors. For normal traffic conditions, detector data was collected through random sampling with the filter that ensured that none of the data belonged to any timestamp where a crash took place on that direction of the route within three hour before or after that time. Afterwards, four classical two-phase traffic flow models: Greenshields, Greenberg, Underwood and Bell-shaped models, were separately built for each detector location and vi each timestamp using the pre-crash dataset. Similarly, the four relationships were generated using the normal traffic condition dataset. The results show that all four classical two-phase traffic flow models are held true for both normal and pre-crash traffic conditions. Downstream detectors and nearest detectors could explain crash data better than upstream detectors. Free flow speed and jam density values of classical models which are depended on model accuracy can explain the variations between normal and hazardous traffic condition