步态是一种极为重要的生物特征,关于步态定量分析的研究具有重大意义。步态定量分析可以应用于多个领域，包括疾病诊断、康复治疗、步态鉴定、体育训练等。本次研究首先总结回顾了目前较为流行的步态定量分析方法及其存在的局限性，设计了一个低成本、无接触、易操作、高效率、适用范围广、分析结果可靠的基于人体三维骨架的步态定量分析系统。为了保证以上方案的顺利实施，本文以国内外现有的相关研究为基础，主要开展了以下工作：

（1）针对单视角单人三维骨架估计，本文提出了一种基于局部关节自注意机制的单视角三维骨架估计法。当前的SOTA模型虽然考虑到了全局空间自注意机制，但是忽略了人体关节之间的自然连接和人体活动的特点。因此，针对单视角单人三维骨架估计，本次研究设计了一种全局和局部空间关节自注意机制相结合的新模型。该模型在原始的MixSTE^[24]模型的基础上添加了一个局部空间自注意模块，用于探究不同动作的关节局部相关性以及提升模型的预测效果。在加入了局部空间自注意模块后，部分测试集的平均每个关节的位置误差有了不同程度的减少。

（2）针对多视角多人三维骨架估计，本文针对基于体素的多视角三维骨架估计代表模型VoxelPose^[23]进行了优化。首先，对于原模型的热图采样融合策略，由于原始热图尺寸较小，直接对原始热图进行采样获得的信息有限，可能导致二维到三维的预测效果欠佳。本次研究在原模型的基础上加入了一个上采样模块，将二维热图的尺寸放大到原来的两倍。加入该模块后，平均每个关节的位置误差下降了0.36mm，不同阈值下的精确度都有了一定程度的提升。其中关节位置误差小于25mm的预测精确度提升效果最为显著，大约提高了3%。实验证明该模块的加入改善了模型对于二维热图的采样融合效果，进而提高了三维骨架估计的精确度。其次，原模型对视角变化比较敏感，当训练时的视角和测试时的视角不一致时，预测的精确度会大幅下降。针对这一问题，本次研究设计了一种基于相机参数对原始热图进行微调的策略：在原模型的基础上加入了一个权重模块和一个动态卷积模块，权重模块将输入视角的相机参数映射为一组卷积参数，动态卷积模块再利用权重模块的输出对原热图进行卷积操作。与利用原模型进行随机视角训练的实验结果相比，精确度有了大幅度的提升。其中关节位置误差小于25mm的预测精确度提高了大约12.5%，平均每个关节的位置误差减小了2.45mm。实验证明本次优化方案在维持与固定视角相近的精确度的基础上，又大大提升了模型对于视角变换的自适应能力。

（3）本文基于目前较为流行的步态定量分析方法以及上述研究结果，设计了一个具有多重优势的基于人体三维骨架的步态定量分析系统。该系统支持单视角视频和多视角视频两种分析途径，涉及多个时空参数和运动学参数。为了获得更加科学可靠的步态定量分析结果，本次研究采集了国内一千名志愿者的正常步态，分别统计了各年龄段的多个步态参数的最大值、最小值和均值作为分析的参考值，并绘制了覆盖整个步态周期的参考值变化曲线，为进行科学的步态分析提供了坚实的基础。

Gait is an extremely important biological feature, and research on quantitative analysis of gait is of great significance. Gait quantitative analysis can be applied in multiple fields, including disease diagnosis, rehabilitation treatment, gait identification, sports training, etc. This study first summarized and reviewed the currently popular gait quantitative analysis methods and their limitations. A low-cost, non-contact, easy to operate, efficient, widely applicable, and reliable gait quantitative analysis system based on the human three-dimensional skeleton was designed. In order to ensure the smooth implementation of the above plan, this article is based on existing relevant research at home and abroad, and mainly carries out the following work:

(1) This paper proposes a single view 3D skeleton estimation method based on local joint self attention mechanism for single view 3D skeleton estimation. Although the current SOTA model takes into account the global spatial self attention mechanism, it ignores the natural connections between human joints and the characteristics of human activities. Therefore, a new model combining global and local spatial joint self attention mechanisms was designed for single view single person 3D skeleton estimation in this study. This model adds a local spatial self attention module on the basis of the original MixSTE [24] model to explore the local correlation of joints in different actions and improve the prediction performance of the model. After adding the local spatial self attention module, the average position error of each joint in some test sets has been reduced to varying degrees.

(2) For multi view and multi person 3D skeleton estimation, this paper optimizes the representative model VoxelPose [23] based on voxel for multi view 3D skeleton estimation. Firstly, for the heat map sampling and fusion strategy of the original model, due to the small size of the original heat map, the information obtained by directly sampling the original heat map is limited, which may lead to poor prediction performance from 2D to 3D. This study added an up-sampling module on the basis of the original model, which enlarged the size of the two-dimensional heat map to twice the original size. After joining this module, the average position error of each joint decreased by 0.36mm, and the accuracy was improved to a certain extent under different thresholds. The prediction accuracy improvement effect is most significant when the joint position error is less than 25mm, with an increase of approximately 3%. Experimental results have shown that the addition of this module improves the sampling and fusion performance of the model for two-dimensional heat maps, thereby enhancing the accuracy of three-dimensional skeleton estimation. Secondly, the original model is sensitive to changes in perspective, and the accuracy of prediction will significantly decrease when the perspective used during training and testing is inconsistent. To address this issue, this study proposes a strategy for fine-tuning the original heatmap based on camera parameters: a weight module and a dynamic convolution module are added to the original model. The weight module maps the camera parameters of the input perspective to a set of convolution parameters, and the dynamic convolution module then uses the output of the weight module to perform convolution operations on the original heatmap. Compared with the experimental results of using the original model for random perspective training, the accuracy has been significantly improved. The prediction accuracy for joint position errors less than 25mm increased by approximately 12.5%, and the average position error per joint decreased by 2.45mm. Experimental results have shown that this optimization scheme significantly improves the model's adaptive ability to perspective changes while maintaining accuracy similar to fixed perspectives.

(3) This article designs a gait quantitative analysis system based on the current popular gait quantitative analysis methods and the above research results, which has multiple advantages and is based on the three-dimensional skeleton of the human body. The system supports two analysis methods for single view and multi view videos, involving multiple spatio-temporal and kinematic parameters. In order to obtain more scientific and reliable quantitative analysis results of gait, this study collected the normal gait of 1000 volunteers in China, and calculated the maximum, minimum, and mean values of multiple gait parameters for each age group as reference values for analysis. A reference value change curve covering the entire gait cycle was drawn, providing a solid foundation for scientific gait analysis.