Lightweight Detection Algorithm for Dense Pedestrian in Airport Terminals

To address the issues of severe occlusion, large variations in target scale, and high computational complexity encountered by manned robots in autonomous navigation tasks within dense crowds in airport terminal scenarios, this paper proposes a lightweight improved algorithm based on YOLOv11n. First, an SCCA module with a hierarchical receptive field mechanism is designed to replace the original C3K2 module, integrating local fine-grained features with global contextual information to enhance multi-scale target feature extraction capabilities. Second, the CGA mechanism is adopted to optimize the C2PSA module, enhancing the capture of detailed semantic features through feature segmentation and cross-head information interaction. Finally, a lightweight detection head LSDDH is designed, introducing DEConv based on the concept of shared convolutions to significantly reduce the number of parameters. Experiments on the self-built terminal building dataset show that the improved algorithm achieves improvements of 1.37% and 2.29% in AP50 and AP50:95 metrics, respectively, compared to YOLOv11n, with a 15.38% reduction in the number of parameters. Generalization experiments on the WiderPerson dataset show corresponding improvements of 1.31% and 1.51%, validating the algorithm's robustness. The algorithm effectively reduces model complexity while improving detection performance, providing a solution for low-power, high-precision pedestrian detection in complex scenarios such as airport terminals.