【摘要】 随着现代化办公方式的普及，长时间保持坐姿已成为常态。然而，现有坐姿识别算法主要依赖人体姿态估计技术，在复杂场景下，受遮挡导致的特征丢失和类内混叠问题影响，识别精度有限。为提升识别性能，提出了双感知混合专家注意力模块（Dual Perceptual Mixture of Attention Heads, DPMoA），结合混合专家模型与多头注意力机制，在路由网络中引入噪声项，以动态激活不同注意力专家头，从而增强泛化能力。具体而言，空间感知交互机制（Spatial Perception Interaction, SPI）与内容感知交互机制（Content Perception Interaction, CPI）分别优化全局信息捕获和局部遮挡处理。SPI通过动态激活关键专家头，实现特征提取与全局交互；CPI针对遮挡问题构建局部特征模型，提升对局部特征的感知能力。实验结果表明，DPMoA在减少模型参数的同时，有效提升了复杂场景下的识别精度。更多还原

【Abstract】 With the widespread adoption of modern office environments, prolonged sitting has become the norm.However, existing posture recognition algorithms primarily rely on human pose estimation techniques, which suffer from feature loss due to occlusion and intra-class ambiguity in complex scenarios, limiting recognition accuracy. To enhance recognition performance, a Dual Perceptual Mixture of Attention Heads(DPMoA) module is proposed, which integrates the mixture of experts model with the multi-head attention mechanism. By introducing a noise term into the routing network,DPMoA dynamically activates different attention heads to improve generalization capability. Specifically, we design the Spatial Perception Interaction(SPI) and Content Perception Interaction(CPI) mechanisms to optimize global information capture and local occlusion handling, respectively. SPI dynamically activates key expert heads to facilitate feature extraction and global interaction, while CPI constructs a local feature model to enhance occlusion-aware perception.Experimental results demonstrate that DPMoA effectively improves recognition accuracy in complex scenarios while reducing model parameters.更多还原