Real-time vehicle queue length prediction using YOLOv11 and machine - learning techniques

Vehicle queue length is an important parameter in traffic congestion, which provides valuable information for effective traffic management and control. The paper proposes real-time vehicle queue length prediction using a combination of computer vision techniques, specifically YOLOv11, and machine learning-based prediction. YOLOv11, a state-of-the-art object detection model, is adopted to detect vehicles captured from a closed-circuit television camera. Based on the detected vehicles, the queue length is computed and the machine learning techniques perform short-term prediction. Several machine learning techniques, including the recurrent neural network, long short-term memory (LSTM) networks, Random Forest, and eXtreme Gradient Boosting, are evaluated. The effects of the time step and introduction of weather data are also considered. The algorithms are implemented on an embedded device to evaluate their real-time performance. The experimental results show that the best prediction is achieved by the LSTM with a time step of 128 using multiple features (queue length and weather data). An evaluation across three signalized intersections (nine cameras) yields a mean absolute error of 7.10, 6.96, and 4.32 m for Intersections A, B, and C, respectively, and a root mean square error of 8.79, 8.60, and 5.27 m, respectively. These results are lower than those reported in related existing work. Furthermore, the algorithm runs with training and prediction times of 3.95 min and 7.55 ms, respectively which is sufficiently fast compared with the time interval of 10 min used in this real-time application.