The design of a content-aware urban traffic management and monitoring system based on a swarm of drones is a challenging problem due to real-time data processing, multi-agent coordination, and decision-making under dynamic conditions. The complexity of this design problem increases with the number of options and the number of cues that traffic controllers must consider when making a decision. The research presented in this thesis addresses this challenge by exploring the potential of swarm intelligence and reinforcement learning-based controllers as design tools for developing adaptive, content-aware decision-making mechanisms for urban traffic management and monitoring using a swarm of drones. The main objective is to design distributed control mechanisms that support the emergence of robust, scalable, and effective autonomous traffic monitoring and management, while proposing and evaluating alternative solutions compared to previous research. In particular, we propose alternative approaches to elements such as the use of predefined correlations between traffic dynamics and driving characteristics, as well as the use of fixed monitoring mechanisms—such as cameras or radar sensors—combined with fixed control strategies.
With regard to traffic monitoring, this thesis focuses on the evaluation of swarms of drones for traffic data collection. With regard to traffic management, this thesis focuses on the evaluation of two traffic control methodologies applied to three different urban scenarios. The control methodologies are: the centralized Simultaneous Perturbation Stochastic Approximation (SPSA) and the decentralized Deep Q-Network (DQN). Briefly, in the former method, the agents operating on traffic collectively evaluate the effects of their actions on traffic quality to update their decision-making policy. The latter method requires agents managing traffic to collectively select the best action from among several alternatives based on the estimated quality of those actions.
In the centralized SPSA controller, we investigate both single- and multi-agent intersection maps using a cell transition traffic model (CTM) to simulate traffic. The experimental results indicate that the proposed drone-based traffic controller can successfully mitigate traffic jams and coordinate the swarm of drones, despite constraints on individual exploration and limited communications at a single urban intersection. However, in multi-urban intersections, reducing traffic in one direction inevitably creates traffic problems in other parts of the road network due to a lack of coordination among agents managing traffic. The study also highlights the trade-off between adjacent intersections in the traffic pattern, where limiting communication can improve the swarm’s adaptability to changes. The evolved optimization controllers outperform traditional mechanisms—such as fixed traffic timing—in terms of accuracy, speed of convergence, varying vehicle input rates, and variations in the traffic model.
In the second scenario, we first evaluate the performance of a centralized SPSA controller under realistic traffic dynamics using the Vissim simulator to simulate road traffic, rather than the CTM model. The aim is to evaluate the controller’s robustness against external disturbances arising from heterogeneous driving behavior. Next, we design and analyze a decentralized DQN controller for a multi-agent intersection map using the Vissim simulator. In this scenario, we extend the investigation to a more complex task in which drones must collectively transfer the collected data between intersections, and each agent selects its timing action based on both its own traffic data and information shared by other agents. The evolved drone-based DQN controllers enable the traffic units to learn about perceptual driving behavior and external disturbances from the Vissim traffic model. The results demonstrate the robustness and scalability of the evolved strategy under various conditions, including different initial conditions and driving behaviors. The swarm of drones also demonstrates adaptability to changes in traffic conditions, although this adaptability depends on the specific nature of the traffic dynamics and the information shared by other agents in the decision-making process.
In the final scenario, we investigate multi-intersection traffic control on a simulated real map of the Belgian city of Namur, using a decentralized Deep Q-Network controller for traffic management and a swarm of drones for monitoring. The traffic model incorporates realistic features such as vehicle acceleration and deceleration, lane-changing behavior, and heterogeneous vehicle types. External disturbances, including parking events and pedestrian crossings, are also considered. The results demonstrate that the proposed autonomous traffic management and monitoring system is robust against disturbances and scalable to real urban areas. The study also highlights the effect of using a swarm of drones for data monitoring, compared to the same control strategy employing fixed cameras, in terms of convergence performance.
In conclusion, this thesis contributes to the field of intelligent traffic management and monitoring by evaluating the potential of a swarm of drones as an effective tool for designing adaptive and coordinated traffic control and monitoring mechanisms. In the scenario under consideration, the results indicate that drone-based traffic controllers outperform conventional fixed-time traffic control and fixed-camera monitoring systems in terms of robustness, scalability, and adaptability. The findings of this research have implications for developing more resilient and autonomous drone-based traffic management systems capable of making informed collective decisions in complex and dynamic urban environments.
