一种基于特征融合的轻量级姿态估计算法  

A Lightweight Pose Estimation Algorithm Based on Feature Fusion

作　　者：张伟杰 叶锋[1] 李昊珑 ZHANG Weijie;YE Feng;Li Haolong(College of Computer and Cyber Security,Fujian Normal University,Fuzhou 350117,China;College of Computer and Data Science,Fuzhou University,Fuzhou 350116,China)

机构地区：[1]福建师范大学计算机与网络空间安全学院,福建福州350117 [2]福州大学计算机与大数据学院,福建福州350116

出　　处：《福建师范大学学报(自然科学版)》2025年第3期19-26,共8页Journal of Fujian Normal University:Natural Science Edition

基　　金：福建省科技创新战略研究联合计划项目(2023R0156)。

摘　　要：在人体姿态估计任务中,多数工作主要关注模型的准确性,忽略了与效率相关的因素,例如模型大小、参数量、推理时间等,而这些指标对实际应用至关重要。针对上述问题,提出一种基于YOLOv8-pose算法的轻量化人体姿态估计算法。算法加入了跨尺度特征融合模块(cross-scale feature fusion module,CCFM),增强模型对于小尺度对象的检测能力,有效结合细节特征和上下文信息,并降低了模型的参数量。同时使用SENetV2替换了YOLOv8-pose中C2f模块的卷积结构,加强了模型在全局角度的考虑,提高了模型的预测精度。使用MPDIoU损失函数提高了模型在训练过程中对边界框误差的计算和模型的推理精度,在mAP50∶95指标上相比原模型提高了1.3%。In human pose estimation tasks,most existing studies primarily focus on the accuracy of models while overlooking efficiency-related factors such as model size,parameter count,and inference time.However,these metrics are crucial for practical applications.To address this issue,this paper proposes a lightweight human pose estimation algorithm based on the YOLOv8-pose algorithm.The algorithm incorporates a cross-scale feature fusion module(CCFM)to enhance the model's adaptability to scale variations and its detection capability for small-scale objects.By effectively combining detailed features and contextual information,the model's overall performance is improved,and its parameter count is reduced.Additionally,SENetV2 is used to replace the convolutional structure in the C2f module of YOLOv8-pose,strengthening the model's global consideration and improving its prediction accuracy.The adoption of the MPDIoU loss function further improves the model's calculation of bounding box errors during training,thereby boosting inference accuracy.The proposed approach achieves a 1.3% improvement in mAP50∶95 compared to the original model.

关 键 词：姿态估计 YOLOv8 轻量化 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]