Abstract
Federated Learning (FL) has emerged as a promising paradigm for decentralised model training in Intelligent Transportation Systems (ITS), offering improved data privacy and reduced communication overhead. Despite growing adoption, the influence of model initialisation strategies on convergence and detection performance in federated vision-based systems remains insufficiently studied. This paper investigates the impact of domain-specific pretraining versus generic weight initialisation on federated object detection under controlled conditions. Using the KITTI dataset, a YOLOv8n-based detection model is evaluated within a horizontal federated learning framework comprising five simulated clients and the FedAvg aggregation strategy. Two experimental phases are considered: federated learning initialised with centrally pretrained domain-specific weights and federated learning initialised with stock model weights. Performance is assessed using precision, recall, and mean Average Precision (mAP). Experimental results show that domain-specific pretraining enables stable convergence and competitive detection accuracy, whereas stock initialisation fails to converge meaningfully under the same training conditions. These findings indicate that model initialisation plays a decisive role in federated object detection performance under limited and partitioned client data.