Can we realize seamless traffic safety at smart intersections by predicting and preventing impending crashes?

Abstract

Intersections are frequently identified as crash hotspots for roadways in major cities, leading to significant human casualties. We propose crash likelihood prediction as an effective strategy to proactively prevent intersection crashes. So far, no reliable models have been developed for intersections that effectively account for the variation in crash types and the cyclical nature of Signal Phasing and Timing (SPaT) and traffic flow. Moreover, the limited research available has primarily relied on sampling techniques to address data imbalance, without exploring alternative solutions. We develop an anomaly detection framework by integrating Generative Adversarial Networks (GANs) and Transformers to predict the likelihood of cycle-level crashes at intersections. The model is built using high-resolution event data extracted from Automated Traffic Signal Performance Measures (ATSPM), including SPaT and traffic flow insights from 11 intersections in Seminole County, Florida. Our framework demonstrates a sensitivity of 76% in predicting crash events using highly imbalanced crash data along with real-world SPaT and traffic data, highlighting its potential for deployment at smart intersections. Overall, the results provide a roadmap for city-wide implementation at smart intersections, with the potential for multiple real-time solutions for impending crashes. These include adjustments in signal timing, driver warnings using various means, and more efficient emergency response, all with major implications for creating more livable and safe cities.

Keywords: Anomaly detection; Crash likelihood prediction; GANs; Proactive safety measures; Smart cities; Smart intersections; Transformers.