Machine learning-based real-time crash risk forecasting for pedestrians

Recent developments in artificial intelligence (AI) have made significant improvements in understanding and enhancing pedestrian safety—a vulnerable road user group that receives less attention than motorized road users do. Specifically, AI-based video analytics have provided insight into facilitating real-time safety at signalized intersections. However, past studies have not fully realized the essence of real-time analysis, which underpins forecasting pedestrian collision likelihood by analyzing how past extreme events influence future risk over sequential intervals. To this end, we combine extreme value theory and machine learning models for real-time pedestrian collision risk forecasting. Traffic conflicts and their associated variables were identified from 288 ​h of video footage obtained from three signalized intersections in Queensland, Australia, via computer vision techniques, including YOLO and DeepSORT, to obtain the post encroachment time for vehicle‒pedestrian interactions. A Bayesian non-stationary peak over threshold (POT) is developed to obtain real-time pedestrian crash risk at the signal cycle level. The performance of the POT model is compared with observed crashes, and the results demonstrate the reasonable accuracy of the model. The estimated pedestrian crash risk at each signal cycle forms contiguous univariate time series data (which serve as ground truth), which are used as input to develop time series machine learning models (recurrent neural networks (RNNs) and long short-term memory (LSTM)). Both of these models forecast pedestrian crash risk, with the RNN model outperforming the competing model and demonstrating that pedestrian crash risk can be reliably estimated 30−33 ​min in advance. • A bi-level framework of extreme value theory and machine learning forecasts pedestrian crash risk. • Safety-critical interactions for pedestrians were measured by post-encroachment time. • A Bayesian peak over threshold model estimates the crash risk at the signal cycle level. • The Recurrent Neural Network (RNN) model forecasts crash risk up to 12 signal cycles.