<p>Multi-source domain adaptive object detection (MSDAOD) is a critical problem in surveillance systems, where distributed cameras capture heterogeneous visual data under diverse conditions. Privacy regulations often prohibit direct data sharing between different cameras, posing significant challenges in building robust detection models. In this paper, we propose a series of frameworks tailored to address MSDAOD under three levels of privacy constraints, ordered by increasing stringency: (i) <i>w/ target</i>, where clients have access to target data; (ii) <i>w/o target</i>, where clients can only access source data while the server can access models trained by clients; and (iii) <i>black-box</i>, where neither clients can access target data nor can the server access model parameters. To solve MSDAOD problems under various constraints, we incorporate domain adaptation and source-only learning techniques at the client side, and design model aggregation strategy and auxiliary server-side mechanisms for privacy-preserving coordination. We conduct extensive experiments on widely used vehicle surveillance datasets, and the results demonstrate that our framework significantly enhances detection performance in privacy-sensitive systems, providing a practical solution for robust cross-domain object detection.</p>