A deep learning-assisted mathematical model for decongestion time prediction at railroad grade crossings

This paper presents a deep learning-assisted framework to estimate the decongestion time at the grade crossing, and its key novelty lies in a differential approach to address the challenge associated with data deficiency of congestion events in grade crossings. A hypothesis of the traffic behavior during the congestion event caused by passing trains is proposed. A deep neural network-based vehicle crowd counting algorithm is developed to estimate the number of vehicles at the normal traffic condition. A running average-based motion detection algorithm is designed to estimate the time of the train passing through the grade crossing. A regression model is then constructed to relate the quantitative information with the decongestion time. In the experiments, 30 congestion events are video-recorded during a period of 200 h with different camera angles at a selected grade crossing, and then studied by the proposed method to learn the congestion pattern and predict the decongestion time, which to the best of our knowledge has not been attempted before. Analysis of the experimental results shows that the vehicle number at the normal traffic flow and the train passing time have significant influences on the traffic decongestion time. The relationship is captured by a quantitative model for rapid prediction. Our study also points out the direction for further improvement of the present development to meet the need for real-world applications.