Safer, more economical and efficient way of traveling is persistently pursed for the development of future traffic systems. Ongoing R&D projects aim at a transition to digital intelligent traffic systems or Smart Cities, in which comprehensive and rich sensor data are applied for improving the safety, traffic efficiency and convenience of traveling, and enable advanced AI-based traffic services for the next generation of traffic systems.
Federated Learning (FL) is a recently proposed distributed learning method, which is well suited for intelligent traffic systems and addresses the machine learning problems: (i) unbalanced local training data in non-IID among devices, (ii) massively distributed devices, (iii) expensive communication. All these features are fulfilled in the domain of intelligent road traffic.
As FL only uploads the locally updated model parameters instead of raw data, the required communication load among traffic agents is usually lower than normal centralized learning methods. Moreover, compared to other distributed learning methods, FL can train learning models from broad users without accessing their private data, e.g., traveling data, raw images, etc. Hence, we believe that FL will bring great benefits in AI-powered traffic systems.