无线电工程2025，Vol.55Issue(8)：1607-1617,11.DOI:10.3969/j.issn.1003-3106.2025.08.007
低光动态场景下的轻量小目标人体姿态估计
A Lightweight Network for Small-object Human Pose Estimation in Low-light Dynamic Conditions
摘要
Abstract
To address the challenges faced by pose estimation networks in low-light environments—such as large model size,difficulty in detecting human targets with small image occupancy due to long-distance shooting,and inaccurate keypoint localization caused by motion blur—a lightweight pose estimation network named LLDS-Pose,based on an improved YOLO11-Pose architecture is proposed.Specifically,Ghost Convolution(GhostConv)is introduced to perform lightweight linear transformations,enhancing feature extraction capability while significantly reducing model parameters.A small-object detection layer is added,and a multi-path fusion strategy is employed to construct a four-scale feature fusion network,thereby improving the network's sensitivity to small-object human targets.Furthermore,a Contextual Transformer Attention(CoTAttention)module is integrated to capture both static and dynamic contextual relationships around local keypoints,effectively enhancing the network's ability to extract fine-grained features from blurred regions.To improve the adaptability and generalization of network under low-light conditions,the Common Objects in Context(COCO)human pose dataset is augmented using techniques such as brightness attenuation,motion blur,and geometric transformations.Experimental results demonstrate that LLDS-Pose achieves 83.7%mAPpose50 and 88.3%mAPperson50 on the COCO dataset,with a parameter count of only 2.3 M and an inference speed of 319.6 frames/s.These results surpass those of mainstream algorithms in both accuracy and efficiency.The superior performance of LLDS-Pose provides a promising and practical solution for human pose estimation tasks in low-light environments.关键词
人体姿态估计/轻量化/低光环境/小目标/动态模糊Key words
human pose estimation/lightweight/low-light environment/small object/motion blur分类
信息技术与安全科学引用本文复制引用
张嘉欣,赵子月,陈猛..低光动态场景下的轻量小目标人体姿态估计[J].无线电工程,2025,55(8):1607-1617,11.基金项目
国家自然科学基金青年科学基金项目(12002115) National Natural Science Foundation of China Youth Science Fund Project(12002115) （12002115）
