High-Precision Hand Pose Estimation Based on Small Object Augmented Feature Pyramid

Hand joint pose assessment is a critical step in the forensic identification of range of motion, holding significant importance in the field of forensic science. To overcome the limitations of traditional methods that rely heavily on professional physicians, this paper proposes an improved hand pose assessment method based on computer vision. Focusing on 21 hand keypoints, the method addresses challenges such as their small size in images, susceptibility to occlusion, and limited feature information by constructing an optimized detection framework. Specifically, a Small Object Augmented Feature Pyramid is designed to alleviate the difficulty of losing small keypoint features; the SPDConv module is introduced to preserve detailed features; and the Omni-Frequency-Spatial Attention Kernel (Omni-FSAK) module is incorporated to achieve efficient multi-scale feature fusion, thereby significantly enhancing the perception capability for blurred, occluded, and tiny keypoints. In terms of detection head design, in addition to the classification branch, bounding box regression branch, and keypoint prediction branch, we further introduce a Localization Quality Calibrator (LQC). This mechanism fully utilizes the outputs from the classification branch and regression distribution information to evaluate the reliability of prediction results. It effectively compensates for the insufficiency of classification confidence in accurately reflecting localization quality, enabling the model to output high-quality detection and keypoint prediction results more robustly under the same network scale. Experiments on the FreiHAND and KeypointsHand datasets verify the effectiveness of the proposed method. Compared with the baseline method, the accuracy is improved by 3.15% and the recall rate is increased by 1.74%, fully validating the feasibility of the improved scheme. 手部关节姿态评估是手部关节活动度司法鉴定中的关键环节，在司法鉴定领域具有重要意义。为克服传统方法对专业医师高度依赖的局限，本文提出了一种基于计算机视觉的改进型手部姿态评估方法。该方法聚焦于手部21个关键点，并针对其在图像中尺寸较小、易被遮挡、特征信息有限等问题，构建了优化的检测框架。具体而言，我们设计了小目标增强特征金字塔，以缓解小关键点特征易丢失的难题；引入SPDConv模块以保留细节特征；并结合Omni-FSAK（Omni-Frequency-Spatial Attention Kernel）模块，实现高效的多尺度特征融合，从而显著提升对模糊、遮挡及细小关键点的感知能力。在检测头设计方面，除分类分支、边界框回归分支和关键点预测分支外，我们进一步引入了位置质量校准器（LQC, Localization Quality Calibrator），充分利用分类分支输出与回归分布信息，对预测结果的可靠性进行评估。该机制有效弥补了分类置信度无法准确反映位置质量的不足，使模型在相同网络规模下能够更稳健地输出高质量的检测与关键点预测结果。实验在FreiHAND和KeypointsHand数据集上验证了所提出方法的有效性，相较基线方法，准确率提升了3.15%，召回率提升了1.74%，充分验证了改进方案的可行性。