复杂场景下的多人人体姿态估计算法

打开文本图片集

关键词： 复杂场景; 多人人体姿态估计; 分组卷积; 空间注意力机制; 轻量化中图分类号： TP391 文献标志码： A 文章编号： 1671-6841（2025）04-0001-07DOI： 10. 13705 / j. issn. 1671-6841. 2024027

Abstract： The cross-obscuration of individuals in complex scenes led to low accuracy and incorrect skeleton connections in existing human pose estimation algorithms. Therefore， a multi-person pose estimation optimization algorithm in complex scenes was proposed. Firstly， the ordinary convolution was replaced with the grouped cascade convolution， which was combined with feature fusion to promote the exchange of information between channels. The accuracy of the algorithm was improved without incurring additional computational costs. Secondly， the spatial attention mechanism was introduced to mine the spatial semantic features related to the human pose estimation task， and the network structure was parallelized to enhance the performance of the algorithm. Finally， the embedding positions of the large convolutional kernel and the attention mechanism were lightweighted to reduce temporal overhead. Compared to the existing bottom-up pose estimation algorithm OpenPifPaf++， the proposed algorithm improved the average accuracy by 0. 8 percentage points on the COCO 2017 dataset. Compared with the OpenPifPaf algorithm， the proposed algorithm improved the average accuracy by 1. 2 percentage points on the CrowdPose dataset， and the corresponding accuracy for complex scenes by 1. 5 percentage points.

Key words： complex scene; multi-person pose estimation; group convolution; spatial attention mechanism; lightweight

0 引言

近年来，深度学习的飞速发展使得基于图像、视频的人体姿态估计技术取得了日新月异的进步。（剩余11237字）