The combination of federated learning and object detection techniques allows for the use of extensive datasets. This approach creates target detection models that are robust and adaptable. It also upholds strict privacy and security protocols. However, the computational resource variance among edge nodes poses significant constraints on the scalability of detection algorithms, thereby impinging on their accuracy and operational efficiency. This discrepancy compels clients, who might otherwise engage complex models, to resort to more simplified algorithms. To navigate this challenge, we introduce a federated learning framework enriched with a knowledge distillation strategy, specifically engineered for the object detection realm. This framework intelligently selects the largest model from the pool of client-contributed models and facilitates model upload and dissemination through a knowledge transfer mechanism orchestrated at the server side. Extensive experimental validations underscore the method’s proficiency in addressing the heterogeneity of detection algorithms within a varied environment, showcasing its viability for real-world application.