自动化与信息工程2025，Vol.46Issue(6)：24-31,8.DOI:10.12475/aie.20250604
融合全局子采样注意力与定向随机失活的电梯场景目标检测模型
An Elevator Scene Object Detection Model Integrating Global Subsampling Attention and Targeted Dropout
摘要
Abstract
Addressing the challenges in elevator scenarios—such as uneven illumination inside the cabin,high-density occlusions,and stringent real-time detection requirements—which make it difficult for object detection models to balance computational efficiency with generalization capability,a GSAOD model framework integrating Global Subsampling Attention(GSA)and Directed Random Dropout is proposed.The GSA module is embedded into the object detection framework,where multi-scale dynamic sub-window sampling is employed to capture key attention regions,preserving spatial awareness and long-range dependencies while reducing the computational complexity of traditional self-attention to linear level.During model training,a directed random dropout mechanism is introduced and specifically applied to attention layers and the feature pyramid network to mitigate overfitting risks.This framework is implemented on the classic YOLO model and validated through experiments on the COCO dataset and a self-built elevator scene dataset.Experimental results show that the proposed model maintains consistent detection performance in general scenarios;in elevator scenarios,the YOLOl_GD model improves mAP@50 and mAP@50:95 by 2.6 and 2.2 respectively compared to the baseline YOLOl model,effectively enhancing robustness in complex industrial environments such as elevators while ensuring real-time detection speed.关键词
目标检测/电梯场景/注意力机制/全局子采样注意力/定向随机失活Key words
object detection/elevator scene/attention mechanism/global subsampling attention/targeted dropout分类
信息技术与安全科学引用本文复制引用
吴宾,于滨,曹健波,朱益霞,方建生,陈再励..融合全局子采样注意力与定向随机失活的电梯场景目标检测模型[J].自动化与信息工程,2025,46(6):24-31,8.基金项目
广东省科学院青年人才专项(2024GDASQNRC-0321) （2024GDASQNRC-0321）
广东省科学院发展专项资金项目(2024GDASZH-2024010102). （2024GDASZH-2024010102）
