| Popis: |
Queue Length Estimation (QLE) at signalized intersections is vital for enhancing urban mobility and managing congestion effectively. This study introduces a novel QLE framework that diverges from traditional methods, which rely on sophisticated sensor settings and shockwave models, by focusing on the kinetic features of vehicles captured at the stopline by sensors. Developed systematically in four stages, the framework begins with a binary classification of vehicles as queuing or non-queuing. This is followed by a lane-based analysis of sequential vehicles to examine following patterns, refining the initial classifications through a label modification strategy, and culminating in the construction of queue sequences for accurate length calculation. A conventional deep learning (DL) method, Multi-Layer Perceptron (MLP), alongside a spatiotemporal approach, Convolutional Neural Network-Long Short-Term Memory-Attention (CNN-LSTM-Attention, C-L-A) were employed for the initial binary classification. Empirical results show the MLP model outperforms C-L-A in this task. Additionally, integrating an Adaptive Neuro-Fuzzy Inference System (ANFIS) with sequential modification enhances the label modification process, facilitating accurate queue sequence generation. Refined vehicle labels are then used for queue length calculation, using a proposed Non-queuing Tolerance Technique (NQTT) and Longest Continuous Queuing Sequence (LCQS). The framework outperforms conventional methods under both oversaturated and undersaturated conditions, with statistically significant improvements validated by one-sided t-tests. Furthermore, in accident-specific scenarios such as rear-end collisions (REC), the framework demonstrates robust performance, effectively handling complexities like increased headway and lane-changing behaviors. This integration of DL and traditional traffic measurement techniques highlights the system’s adaptability in improving QLE accuracy across diverse traffic conditions. |
