Abstract: In the field of human pose estimation, heatmap-based methods suffer from the problems of quantization error, high computational complexity, and the need to post-process the heatmap as well. To address the above problems, a lightweight human pose estimation model based on fused feature state space model was proposed, namely Lite-SimCC, using the SimCC method of coordinate regression as a baseline. ShuffleNetV2 was adopted as the backbone network to replace the original High-Resolution Net (HRNet), which simplified to a single-branch form and realizes the model lightweighting. To mitigate the loss of accuracy, a large kernel convolution was introduced to extract global feature information. A state space model with fused features was further designed to handle both local and full long sequence features to enhance the characterization of key points. Finally, a soft-label based loss function was proposed to replace the traditional one-hot loss calculation. Experimental results show that compared with the baseline method SimCC, Lite-SimCC has the parameters decreased by 87.1%, and the Average Precision (AP) is improved by 1% to 71.8% on the COCO2017 test set. And the experiments on the MPII dataset also prove that Lite-SimCC also effectively reduces the number of parameters of the model based on the guaranteed detection accuracy.

Key words: human pose estimation, coordinate regression, state space model, lightweight, soft-label

摘要： 人体姿态估计领域中，基于热图的方法存在量化误差、计算复杂度高以及还需对热图进行后处理等问题。针对上述问题，以坐标回归的SimCC方法为基线，提出一种基于融合特征状态空间模型的轻量化人体姿态估计方法Lite-SimCC，采用ShuffleNetV2作为骨干网络，替代原有的HRNet（High-Resolution Net），简化为单分支形式，并实现模型轻量化。为了缓解准确率的损失，引入大核卷积提取全局特征信息。进一步设计了融合特征的状态空间模型，用于处理局部和全局长序列特征，增强关键点的表征能力。最后提出一种基于软标签的损失函数，替代传统的one-hot损失计算方式。实验结果表明，与基线方法SimCC相比，Lite-SimCC的参数量少了87.1%，在COCO2017测试集上平均精度(AP)提升了1%达到71.8%，在MPII数据集上的实验也验证了Lite-SimCC在保证检测精度的基础上还有效降低了模型的参数量。

关键词: 人体姿态估计, 坐标回归, 状态空间模型, 轻量化, 软标签