Decentralized Traffic Management Through a Hybrid Incremental and Federated Learning Approach

Abstract

Urban traffic is one of the most important issues for smart cities, and real-time management is critical for improving mobility and reducing congestion. Traditionally, classification methods based on machine learning need frequent retraining, which is inefficient for adaptive traffic management systems. This work introduces a novel hybrid approach that combines Incremental Learning (IL) with Federated Learning (FL) techniques to support continuous model adaptation without centralizing data or restarting the training process from scratch. The proposed approach employs Convolutional Neural Networks (CNNs) to classify traffic conditions from junction camera feeds and includes a specific mechanism to mitigate the Catastrophic Forgetting (CF) issue, a common drawback in IL. This solution enhances model performance in a decentralized way while protecting data privacy and encouraging knowledge sharing across distributed nodes. Experimental results on a publicly available dataset reveal that this approach significantly improves dynamic traffic management, achieving over 96% validation accuracy throughout the IL process for multiple clients, with minimal loss and no need to centralize data. This work sets the basis for a more effective and secure traffic management infrastructure in smart cities.