Enhancing Object Detection in IoV: A Federated Semi-Supervised Learning Approach With Data Assessment

Integrating Connected and Autonomous Vehicles (CAVs) with federated learning (FL) has garnered widespread attention in recent years, particularly in object detection. However, within the Internet of Vehicle (IoV) context, employing FL to handle complex visual tasks faces several challenges, such as difficulties in obtaining labeled data, heterogeneity in user data across vehicles, and limitations in timely assessing user contributions. To address these challenges, we propose a federated semi-supervised learning architecture for object detection, accompanied by a contribution evaluation and aggregation method based on heterogeneous data. Specifically, we designed a federated semi-supervised training process for the IoV, utilizing an object detection framework based on Faster R-CNN and a teacher–student architecture. To demonstrate its effectiveness, we conducted a communication feasibility analysis using real-world vehicular network data and an analysis of the algorithm’s convergence properties. Additionally, we developed a data-based user contribution assessment and aggregation framework to evaluate the distribution and quality of data from vehicle users to aid the FL center. Finally, simulation results show that the proposed federated semi-supervised algorithm can effectively train and converge to a model that outperforms traditional FL. Ablation experiments further validate the efficacy of the data-based assessment method.