基于自适应跨维加权网络的轻量型人体姿态检测

Lightweight high-resolution network based on adaptive cross-dimensional weighting for human pose estimation

摘要:

人体姿态检测的核心是准确检测人体关键点，由于高分辨率网络存在着一定局限性，对此，提出一种自适应跨维加权高分辨率网络；针对网络跨维信息交互不足的问题，采用跨维分离卷积方法提取信息，实现了在空间和通道之间有效的信息交换；针对关键点定位不精确的问题，采用自适应上下文建模方法，通过自适应变换和输入特征的空间加权，增强了网络捕捉复杂空间关系的能力，使得网络能够提取多尺度上下文信息并建立跨维度依赖关系，从而在不增加计算复杂度的情况下提高了准确性；此外，还引入了坐标注意力机制融合来自不同分支和规模的特征，使检测准确性得到进一步提升；经COCO和MPII数据集实验测试，与主流轻量型网络相比，自适应跨维加权高分辨率网络性能更好，兼顾了效率与精度。

Abstract:

Human pose detection fundamentally involves the precise localization of key anatomical points on the human body, which is pivotal for a wide range of visual perception tasks. Despite the notable advancements in high-resolution networks, they exhibit inherent limitations that hinder their performance. To address these challenges, we introduce the Adaptive Cross-Dimensional Weighting High-Resolution Network (ACW-HRNet), a novel architecture designed to enhance pose detection accuracy. Specifically, to mitigate the problem of inadequate interaction between cross-dimensional information, we propose the adoption of cross-dimensional split convolution, a technique that facilitates the efficient exchange of information between spatial and channel dimensions. To further improve the precision of key point localization, we integrate Adaptive Context Modeling (ACM), which augments the network's capacity to capture complex spatial relationships through adaptive transformations and spatial weighting of the input features. This approach enables the network to extract rich, multi-scale contextual information while simultaneously establishing cross-dimensional dependencies, resulting in a marked improvement in accuracy without incurring additional computational overhead. Moreover, we incorporate a coordinate attention mechanism that facilitates the fusion of multi-branch and multi-scale features, further enhancing detection accuracy. Empirical evaluations conducted on the COCO and MPII datasets demonstrate that ACW-HRNet significantly outperforms leading lightweight networks, achieving a harmonious balance between computational efficiency and detection accuracy.