Intelligent path planning using deep learning and reinforcement learning in transportation systems

Nowadays, transportation is one of the major challenges faced by the public, as many people commute daily between their homes and workplaces. To tackle this issue, individuals are encouraged to use public transportation. At the same time, the development of Intelligent Transportation Systems (ITS) is becoming increasingly important in modern society. As a result, people are more willing to adopt transportation systems that are smart, safe, and environmentally friendly. Modern vehicles increasingly incorporate advanced technologies such as object detection, reinforcement learning (RL), deep learning (DL), and machine learning (ML) to improve safety, efficiency, and traffic management. A significant research challenge in autonomous vehicles is robust object detection under various environmental conditions, such as fog, rain, and low-light scenarios. In this study, a YOLO (You Only Look Once)–based object detection model is developed to identify vehicles, pedestrian areas or crosswalks, traffic lights, and road signs on roads. The model is evaluated on images captured in both foggy and sunny conditions, demonstrating its ability to maintain high detection accuracy despite reduced visibility or image noise. This work contributes to enhancing real-time perception in autonomous systems, which is critical for safe navigation and traffic safety. Another important problem addressed in this research is optimal path planning in urban traffic networks, which is essential for reducing congestion and minimizing travel time. A multi-input convolutional neural network (CNN) is employed to process map data from the SUMO (Simulation of Urban MObility) environment, including features such as road edges, traffic light locations, and vehicle distributions. By integrating these features, the CNN is able to determine optimal routes dynamically, considering real-world factors such as traffic density and signal placement. This approach improves the reliability of autonomous navigation in complex urban scenarios. Finally ,the study uses reinforcement learning (RL) to explore dynamic routing. On a simulated metropolitan map of Bari, Italy, a Q-Learning framework is used to determine the shortest and most efficient paths while accounting for potential dead-end routes, traffic lights, vehicle density, and road topology. Through iterative updates of the Q-table, the RL agent learns optimal routes that balance avoiding traffic and minimizing travel time. This approach shows how RL may be used practically for autonomous vehicle routing and adaptive traffic management. Overall, this research combines YOLO-based object detection, CNN-based path planning, and RL-based dynamic routing to provide an integrated framework for autonomous transportation. The proposed methods contribute to enhancing safety, reducing congestion, and improving efficiency in urban traffic systems, offering both scientific insights and practical solutions for modern ITS applications.