基于PCNet的人体姿态估计方法

PCNet:Human Pose Estimation With Pose Correction In Complex Scenes

摘要:

人体姿态估计是计算机视觉、模式识别领域的重要研究问题,在人机交互领域有重要应用。人体姿态估计用于将视频图像中的人体骨骼姿态进行检测识别,研究面向闸机场景下人群拥挤、遮挡严重的复杂场景下人体姿态估计方法,研究工作具有挑战性。首先,针对拥挤、遮挡严重的复杂场景下人体姿态估计任务,我们提出基于姿态矫正的人体姿态估计网络(PCNet,Pose Correction Network),设计了一种融合全局和局部信息的 Transformer 特征编码模块,并将其引入到模型特征提取骨干网络中提升精度表现。第二阶段对预测的关键点位置进行矫正,提出基于时空注意力机制的级联结构的姿态矫正模块,修正因遮挡、小尺度目标等引起的误差较大的关键点。提出的人体姿态估计方法在COCO数据集和CrowdPose数据集上进行实验,实验结果表示,本文提出的模型效果在精度和鲁棒性上均得到了提升,证明了本文所提出的人体姿态估计方法的有效性。

Abstract:

Human pose estimation is a significant research problem in the fields of computer vision and pattern recognition, with crucial applications in human-computer interaction. It involves detecting and recognizing human skeletal poses in video images, particularly in complex scenarios with severe crowd congestion and occlusion, posing considerable challenges. Firstly, addressing the task of human pose estimation in complex scenes plagued by congestion and severe occlusion, we propose a Pose Correction Network(PCNet), introducing a fusion of global and local information in a Transformer feature encoding module. This module is integrated into the backbone network for feature extraction, thereby enhancing accuracy. In the subsequent stage, corrective measures are applied to predicted keypoint positions. We propose a hierarchical posture correction module based on spatiotemporal attention mechanisms, aimed at rectifying significant errors in keypoint localization caused by occlusion and small-scale targets. The effectiveness of the proposed human pose estimation method is evaluated on the COCO and Crowd Pose datasets. Experimental results demonstrate improvements in both accuracy and robustness, validating the efficacy of the proposed approach.