Real-Time Traffic Data Analysis and Deep Learning-based Traffic Volume Classification for Congestion Mitigation at Urban Intersections

Abstract

Managing traffic at intersections in crowded cities is highly dependent on understanding crowding level, which can be assessed through factors such as traffic volume, vehicle count, and signal timing and control. Traditional Intelligent Traffic Systems (ITS) methods, including detector loops, GPS, and camera-based solutions, often present complexities and high costs. This study proposes an alternative approach using Acoustic Traffic Monitoring(ATM) technology to detect abnormalities traffic patterns. A new model was developed to detect traffic volume as crowded (abnormal flow) or non-crowded (normal flow) based on an acoustic dataset collected using Raspberry Pi. The collected data underwent analysis through signal processing techniques, followed by detection using machine learning models: Support Vector Machine (SVM), Random Forest (RF), Artificial Neural Network (ANN), Long Short-Term Memory (LSTM). Evaluating tow datasets, RTD and IDMT-Traffic, across four frame sizes, the results demonstrate the effectiveness of the proposed method. Notably, the LSTM model achieved accuracy of 98.25% on the RTD dataset and 98.69% on the IDMT-Traffic dataset, highlighting it is potential for accurate traffic jam detection.