An advanced deep learning approach to real-time estimation of lane-based queue lengths at a signalized junction

In this study, we develop a real-time and novel estimation method of lane-based queue lengths using two deep learning processes, which include of a Convolutional Neural Network (CNN) into a Long Short-Term Memory (LSTM). This approach not only outperforms the recently developed real-time estimation of lane-based queue lengths but also captures the spatiotemporal attributes of traffic. There are three primary challenges to design a deep learning based queue estimation model. First, the CNN and the LSTM are integrated to estimate lane-based queue lengths minimizing accumulative counting errors. Furthermore, short-term arrival patterns and long-term traffic demand trends are captured by the LSTM to improve the accuracy of estimates of cycle-based proportional lane-uses. In addition, imaged second-based occupancy rates and impulse memories are used to identify whether vehicular queues are remained at the end of each cycle by using the CNN. In numerical examples and case study, the integrated CNN LSTM method shows excellent performance to estimate queue lengths in individual lanes in seconds compared to the other approaches applied in this paper. This work paves the way for the applicability of the deep learning to estimate traffic quantities in real-time for lane-based adaptive traffic control systems (ATCS). Furthermore, we will introduce offset in a signal plan and lane-based turning proportion on the proposed framework to explain vehicular spillbacks in an individual lane and a grid lock for pursuing coordinated traffic movements along arterials and in signalized urban networks.