Dayoub, FerasLiu, LingqiaoShi, Xiangyu2025-08-082025-08-082025https://hdl.handle.net/2440/146662Object detection is a critical capability for mobile robots navigating diverse and dynamic environments. However, standard methods struggle with domain shifts and real-time adaptability, limiting their deployment in real-world scenarios. To address these challenges, this thesis proposes two novel approaches: an Online Source-Free Domain Adaptation (O-SFDA) framework leveraging unsupervised data acquisition, and an Open-Vocabulary Object Detection (OVOD) model for domain adaptation in embodied environments. The first contribution focuses on improving adaptive object detection using OSFDA by introducing an unsupervised data acquisition framework. In mobile robotics, not all captured frames contain valuable information for adaptation, especially in the presence of strong domain shifts. Our method prioritises the most informative unlabeled samples for inclusion in the online training process, significantly enhancing adaptation performance. Empirical evaluation of real-world datasets demonstrates that this approach surpasses existing state-of-the-art (SOTA) techniques, underscoring the viability of selective data acquisition in improving real-time adaptability. The second contribution leverages OVOD as the base model for domain adaptation in indoor environments. To overcome the limitations of existing methods under domain shifts, we propose a Source-Free Domain Adaptation (SFDA) approach that adapts pre-trained models without requiring access to source data. This approach refines pseudo-labels using temporal clustering, incorporates multi-scale threshold fusion for robust adaptation, and employs a Mean Teacher framework enhanced with contrastive learning. Additionally, we introduce the Embodied Domain Adaptation for Object Detection (EDAOD) benchmark to evaluate performance under sequential changes in lighting, layout, and object diversity. Experimental results highlight substantial improvements in zero-shot detection and adaptability to dynamic indoor conditions, demonstrating the effectiveness of our approach. Together, these contributions advance the SOTA in adaptive object detection for mobile robots, addressing critical challenges posed by domain shifts and dynamic environments. By enabling more robust and flexible object detection, this work facilitates the deployment of mobile robots in real-world scenarios, enhancing their ability to navigate and operate effectively in complex, ever-changing conditions.enDomain AdaptationObject DetectionRobot PerceptionThesis