3D object detection based on LiDAR point clouds has been widely explored in autonomous driving. Federated Learning (FL), with its inherent advantages in distributed privacy protection and scalable multi-vehicle collaboration, presents a promising alternative to centralized training. However, its real-world deployment encounters the major challenges: (1) severe LiDAR data heterogeneity across vehicles hinders global model convergence; (2) inherent conflicts between global model consistency and local personalization. To address the challenges, we propose TeamFed, a FL framework for 3D object detection inspired by organizational principles in human teams. By incorporating a multi-leader collaboration paradigm and a personalized sampling strategy, TeamFed enables efficient global model training and effective local personalization tuning. Extensive experiments demonstrate that TeamFed achieves 15.8% and 22.1% performance improvements in global and personalized models respectively under heterogeneous autonomous driving scenarios, approaching or even surpassing the performance of centralized training.