Vulnerable road users' crash hotspot identification on multi-lane arterial roads using estimated exposure and considering context classification. (September 2021)

• Record Type: Journal Article
• Authors: Mahmoud, Nada; Abdel-Aty, Mohamed; Cai, Qing; Zheng, Ou
• Is Part Of: Accident analysis and prevention. Volume 159(2022)
• Publication Date: 2021-09
• Subjects: Bicycle exposure -- Pedestrian exposure -- Machine learning -- Safety performance function -- Statistical model -- Vulnerable road user; Accidents -- Prevention -- Periodicals; Accident Prevention -- Periodicals; Accidents -- Prévention -- Périodiques; 363.106
• DOI: 10.1016/j.aap.2021.106294
• Languages: English
• ISSNs: 0001-4575

Abstract

Highlights: Utilize big data from different sources including ATSPM, Strava, traffic and roadway features, land-use attributes, socio-demographic characteristics, and crash data. Obtain the ground truth data from Closed Circuit Television (CCTV) surveillance camera recorded videos using manual counts and automatic video recognition. Develop machine learning models to estimate hourly vulnerable road users' exposure and bike exposure considering context classification. Identify the hotspots for different context classifications by developing crash prediction models based on big data. Conclude the factors contributing to vulnerable road users' crashes at intersections and bikes crashes along the roadway segments. Abstract: This research develops safety performance functions and identifies the crash hotspots based on estimated vulnerable road users' exposure at intersections and along the roadway segments. The study utilized big data including Automated Traffic Signal Performance Measures (ATSPM) data, crowdsourced data (Strava), Closed Circuit Television (CCTV) surveillance camera videos, crash data, traffic information, roadway features, land use attributes, and socio-demographic characteristics. It comprises an extensive comparison between a wide array of statistical and machine learning models that were developed to estimate pedestrian and bike exposure. The results indicated that the XGBoost approach was the best to estimate vulnerable road users' exposure at intersections as well as bike exposure along the roadway segments. Afterwards, the estimated exposure was utilized as input variables to develop crash prediction models that relate different crash types to potential explanatory variables. Negative Binomial approach was followed to develop crash prediction models to be consistent with the Highway Safety Manual. The results show that the exposure variables (i.e., AADT, bike exposure, and the interaction between them) have significant influences on the two types of crashes (i.e., crashes of vulnerable road users at intersections and bike crashes along the segments). Further, the results indicated that the context classification is significantly related to crashes. Based on the developed models, the PSIs were calculated and the hotspots were identified for the two crash types. It was found that hotspots were more likely to be located near the city of Orlando. Coastal roadways were classified as cold categories regarding bike crashes. Further, C4 roadway segments were found to be significantly related to the increase of vulnerable road users' crashes at intersections and bike crashes along the segments.