Abstract
The detection of mixed road users in intelligent transport systems (ITS) has been based on the use of centralised deep learning mechanisms. These centralised deep learning models requires that a large scale of dataset is aggregated into a central database. This central aggregation raises concerns about the privacy and communication overhead of the system.Federated learning (FL) brings to the table a decentralised alternative that allows dataset to be shared among models in the federation network, however the effect of this decentralisation on the detection accuracy of the federated model remains shallowly explored. This study presents a controlled comparison between a centralised object detection model and a federated learning framework using the YOLOv8n architecture and the KITTI dataset. The centralised model serves as an upper-bound baseline, while the federated model is trained across five simulated clients using the FedAvg aggregation strategy. Performance is evaluated using Precision, Recall, and mean Average Precision (mAP). Experimental results show that the federated model achieves detection performance comparable to the centralised baseline under the evaluated conditions, while preserving data locality. These findings suggest that federated learning can support privacy-aware mixed road-user detection in ITS without a significant loss in accuracy.