Hospital Scene Object Detection with Integrated Lightweight Convolutions and Attention Mechanisms

Object detection, a critical technology in robotic vision, has seen widespread adoption in various applications, especially in robot navigation and obstacle avoidance. Nevertheless, traditional algorithms struggle in dynamic hospital environments, exhibiting high rates of false positives, missed detections, and poor performance on small or occluded targets. The large model sizes further complicate deployment on mobile platforms, limiting real-world utility. To overcome these limitations, we present a hospital-specific pedestrian detection framework based on YOLOv11. Our methodology features three main contributions: First, we replace the C3k2 Bottleneck with GhostConv, significantly reducing model complexity. Second, we integrate GhostConv with ECA attention to form the GhostECA module, improving feature extraction capabilities without increasing parameters. Third, we introduce a dedicated small object detection layer and a CCAR module for enhanced multiscale feature handling. Extensive experiments on the HosSceneDet, KITTI, and SODA10M datasets validate our approach: the optimized S model achieves $94.8 \% \mathrm{mAP}\text{@} 0.5$ on KITTI $(1.7 \%$ improvement), while the N model reduces parameters by $32.1 \%$. These advancements demonstrate superior performance and practicality compared to existing methods.