Framework of Big Data and Deep Learning for Simultaneously Solving Space Allocation and Signal Timing Problem

Efficient operations at intersections are associated with smooth, safe, and sustainable travel at the network level. It is often challenging to prevent congestion at these locations, especially during rush hours, owing to high traffic demand and restraints of road geometry. The primary cause of intersection traffic congestion is large variations and fluctuations in traffic demand during the day. Dynamic lane assignment (DLA) is an intelligent transportation system (ITS) technology that can improve traffic operations at signalized intersections, making better use of available space by optimizing lane allocation based on real-time fluctuating traffic demands. This study used a brute-force optimization model to produce a vast synthetic dataset of turning movements at a four-legged signalized intersection. The model then found the optimal lane assignment and cycle length for each case. The produced data were used to develop a deep learning model to predict the optimal lane assignment and cycle length simultaneously at any four-legged signalized intersection without applying optimization software per se. The deep learning model results developed were compared with the results of the k-nearest neighbors (KNN), random forest, and decision tree algorithms. It was found that the deep learning model was superior, with accuracy and F1 score of 95.26% and 93.56%, respectively. The proposed solution is expected to require insignificant additional resources and services for successful implementation, particularly in developing countries.