Evaluation of vulnerable pedestrian potential risk at intersection based on predicted arrival time using deep learning methods

Abstract

Pedestrian safety is one of the most pressing transportation concerns around the world. Furthermore, vulnerable users suffer from a higher fatality rate than standard pedestrians. However, there is no data to clarify the pedestrian categories in fatal and nonfatal accidents. Currently, real-time detection of vulnerable users using advanced traffic sensors installed at the intersection has great potential to improve traffic safety significantly. This study develops a framework to evaluate the vulnerable pedestrian potential risk based on a new safety indicator, Predicted Pedestrian Arrival Time (PPAT), using deep learning method. With automated computer vision techniques, mobility features of vehicles and pedestrians can be extracted from video and Lidar datasets. With the algorithm based on pedestrian behavior and physical feature, we can identify vulnerable pedestrians. Deep learning methods(GRU, LSTM, Transformer) are performed to predict the time needed to reach a specific location. Finally, systematic warning systems for vulnerable pedestrian potential risk are designed based on the predicted arrival time. Case study at signalize intersection with Lidar dataset and case study non-signalized intersection with video dataset are performed with this framework. The proposed framework shows high performance in evaluating the vulnerable pedestrian potential risk with 86.59\% accuracy.