Aligning Confidence and Localization: An Enhanced DINO Model for Small Object Detection in Robotic Nursing

In care environments such as nursing homes, robots performing tasks (e.g., feeding assistance) must accurately identify and locate target objects to ensure safe and efficient execution. However, real-world applications face challenges such as numerous small objects, complex backgrounds, and severe occlusions, all of which compromise detection performance. To address these challenges, this study proposes an end-to-end object detection algorithm based on an enhanced DINO framework. Transformer-based DINO is adopted as the baseline, leveraging its global modeling capabilities and avoiding the complex pre- and post-processing required by traditional CNN detectors. In addition, an improved Align-Loss is introduced to enhance small-object detection and address misalignment issues within DINO. Furthermore, a GhostConv module is integrated into DINO’s ResNet50 backbone to reduce the computational load of feature extraction and accelerate detection. Finally, multi-scale data augmentation and transfer learning are applied during training to improve detection accuracy and accelerate convergence. To validate the proposed method, experiments were conducted on the augmented MYNursingHome dataset and the COCO dataset. On the MYNursingHome dataset, the proposed approach improved mAP by 3.1% and APs by 2.6% over the DINO baseline, while reducing parameters from 47M to 39.6M and FLOPs from 279G to 243G. On the NVIDIA Jetson Orin Nano Super, inference speed increased from 16.8 FPS to 18.9 FPS (+12.5%). The experimental results demonstrate that the improved DINO detector proposed in this study exhibits a significant advantage in small object detection for nursing scenarios, providing algorithmic support for intelligent and efficient robotic care.

@article{Du2025,
  title     = {Aligning Confidence and Localization: An Enhanced DINO Model for Small Object Detection in Robotic Nursing},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {16},
  number    = {11},
  year      = {2025},
  publisher = {The Science and Information Organization},
  author    = {Yanchen Du and Qingzhuo Yuan and Shengli Luo and XiaoLong Shu and Xu Wang and Yuheng Jiang and Hongliu Yu},
  doi       = {10.14569/IJACSA.2025.0161150},
  url       = {https://doi.org/10.14569/IJACSA.2025.0161150}
}