Effective object detection in mobile robots is challenged by deployment in diverse and unfamiliar environments. Online Source-Free Domain Adaptation (O-SFDA) offers real-time model adaptation using a stream of unlabeled data from a target domain. However, not all captured frames in mobile robotics contain information that is beneficial for adaptation, particularly when there is a strong domain shift. This paper introduces a novel approach to enhance O-SFDA for adaptive object detection in mobile robots via unsupervised data acquisition. Our methodology prioritizes the most informative unlabeled samples for inclusion in the online training process. Empirical evaluation on a real-world dataset reveals that our method outperforms existing state-of-the-art O-SFDA techniques, demonstrating the viability of unsupervised data acquisition for improving adaptive object detection in mobile robots.