In recent years, privacy-preserving multi-source domain adaptation has emerged as a promising direction for collaborative object detection across diverse environments. However, existing studies often assume homogeneous model architectures, which limits their applicability in real-world surveillance systems composed of heterogeneous detectors. To address this issue, we present a novel framework entitled Cross-Domain Federated Surveillance System with Heterogeneous Object Detectors, which enables privacy-preserving and communication-efficient model collaboration without sharing raw data. Our framework adopts federated learning as the foundation, where multiple clients (representing different source domains or detection models) collaboratively improve a shared object detection capability under a unified coordination scheme. On the client side, domain-specific bias is reduced through local adaptation, while on the server side, heterogeneous models are fused to achieve unified representation and enhanced performance. The proposed system bridges the gap between diverse architectures and real-world surveillance scenarios, ensuring both privacy protection and detection accuracy across domains.