An Ensemble Deep Learning Framework for Real-Time Queue Length Estimation at Signalized Intersections

Abstract Optimal traffic control and signal planning can considerably reduce intersection congestion and delays. To accomplish this, it is essential to have accurate knowledge of the current and prospective vehicle queue lengths. This paper proposes a novel ensemble deep learning approach to estimate and predict the cycle-based maximum queue length in real time. The proposed data-driven framework is built on high resolution traffic data obtained from a single loop detector. In contrast to traditional traffic flow theory-based approaches which depend on deterministic models, this methodology relies on learning the traffic patterns and queue length variations from data, allowing it to capture discrepancies within traffic patterns and as a result, produce improved results. The ensemble deep learning method employs two distinct neural networks: a multi-layer feedforward neural network (FNN) that maps the relationship between high resolution data and the cycle-based maximal queue length, and a stacked long–short-term memory (LSTM) model that captures the temporal variation of the queue lengths. The notion of ensemble produced by combining the outputs of both of these neural networks outperform the performance of the base models separately. This algorithm is trained and tested with simulated data and validated with real-world data obtained from an isolated intersection. In addition, the proposed framework is compared to model-based techniques such as the Kalman filter.