多分辨率特征融合的人体下肢关键点检测

打开文本图片集

关键词：人体下肢关键点检测；多分辨率特征融合；注意力机制；预标记中图分类号：TP394.1 文献标识码：Adoi：10.37188/OPE.20253314.2291 CSTR：32169.14.OPE.20253314.2291

Human lower limb keypoint detection based on multi-resolution feature fusion

XIA Xiaohua*，XIANG Haoming， CHEN Jian，FENG Xinmiao，QIU Fabo，WANG Yaoyao （KeyLaboratory ofExpressway Construction Machinery of Shaanxi Province， Chang'an University，Xi’an 71O064，China） *Corresponding author，E-mail：xhxia@chd.edu.cn

Abstract： The existing human keypoint detection models lack attention to high-resolution features，and the dataset used in training has low resolution and large annotation errors.This causes an unstable detection result and poor positioning accuracy in tasks such as gait analysis. A human lower-limb keypoint-detection model based on multi-resolution feature fusion was proposed to address the above issues.High-resolution images were adopted as network inputs，and a fine-tuned MobileNet vl network combined with an attention mechanism was employed to extract global low-resolution features.Preliminary keypoint positions were predicted，after which local high-resolution features were extracted through a paralel shallow network. Subsequently，features at different resolutions were fused via a continuous residual structure and anattention mechanism，thereby improving keypoint-prediction accuracy and effectively alleviating the computational burden imposed by high-resolution images. In this work，a high-resolution，high-precision dataset of human lower-limb keypoints was created through pre-labeling to ensure reliable model training. Model complexity，detection speed，detection accuracy，and detection error were evaluated and compared with other classic and state-of-the-art methods through experiments. The results show that the test detection rate of the proposed model reaches 95.2% ，which is better than the comparison methods of Lightweight-OpenPose，HRNet-W32，HRNet-W48，YOLO-Pose，RTMPose and SimCC，and the detection accuracy is increased by 4.1%-83.6% ，and the FPS is increased by 7.6-13.9. The effectiveness of the proposed model in high-precision human lower limb key point detection is demonstrated.

Key words： human lower limb keypoint detection； multi-resolution feature fusion；attention mechanism; pre labeling

1引言

人体关键点检测是计算机视觉领域中一项基础且具有挑战性的任务，旨在识别并定位图像或视频中人体的关键点位置，比如髋关节和膝关节等身体部位[]，在视频监控、智能交互等领域有着广泛的应用[2-4]，也是人体步态分析的主要方法之-[5]

利用人体关键点检测模型提取关键点坐标，并根据坐标计算相应的步态参数，从而实现人体步态特征的量化评估，可以为全膝关节置换、中风后偏瘫等患者的诊断和治疗提供客观有效的理论支持[67]，让医生快速、准确地掌握患者的术后康复情况，完善治疗方案[8]。（剩余14352字）