基于多尺度视频的抗运动干扰远程光体积描记关键技术研究

This project addresses the pulse measurement inaccuracy of remote photoplethysmography when the subject exhibits strenuous motion during recording. Existing methods suffer from single-scale image processing in obtaining raw measured traces, and from parameter perturbation in the pulse extraction model. Inspired by multi-scale analysis methods in the field of image processing, a novel multi-scale video based pulse signal extraction and fusion method is proposed, which takes full advantages of multi-scale analysis methods. The proposed method will break through the limitation of state-of-the-art methods and is expected to significantly improve measurement accuracy. First, a bi-lateral image pyramid sequence construction method is proposed such that the raw traces can be initially separated and extracted. Second, an optimized projection vector based pulse extraction algorithm is proposed to address the problems of poor adaptive ability and narrow application range of existing methods. Third, a pulse fusion model that considers both intra-scale and inter-scale features is proposed, which further removes the motion artifact. The above three methods correspond to three steps of multi-scale analysis methods, i.e., multi-level partition, extraction, and fusion, which make perfect integration of multi-scale analysis methods and remote photoplethysmography. They also impulse of the intersection and fusion of multiple disciplines, promote the application, and serve the national economy and people's livelihood.

本项目针对现有远程光体积描记方法在被测者剧烈运动时脉搏信号测量精度不高的问题，总结出现有技术在原始测量信号获取方面局限于单尺度图像处理、在信号分离模型方面存在参数摄动问题等缺陷，受图像处理领域多尺度分析方法的启发，提出基于多尺度视频的脉搏信号分层提取与融合的新方法，能够充分发挥多尺度分析方法的优势，有望突破现有方法局限，显著提高脉搏测量精度。第一，提出双向金字塔图像序列构建方法，实现原始测量信号的初始分离和分层提取；第二，提出基于最优投影向量的脉搏信号提取算法，克服现有方法自适应能力差、适用范围窄的问题；第三，提出兼顾层内特征与层间关联的脉搏信号融合模型，充分利用层间关联信息，实现对运动干扰的进一步滤除。以上所提的三个方法分别对应多尺度分析方法中分层、提取与融合三个环节，实现了多尺度分析方法与远程光体积描记的完美结合，有力推动了多领域研究的交叉与融合，促进该技术早日走向应用、服务国计民生。

远程光体积描记（remote photoplethysmography, rPPG）是一种非接触式生理信号测量手段，近年来已成为研究热点，尤其是在新冠流行的当下更显示出其巨大的应用价值。本课题从多尺度图像分析的角度解释rPPG信号提取问题，攻克在二者结合过程中所遇到的物体跟踪、多尺度图像表征构建与脉搏特征提取和融合三方面的关键科学问题。建立了多尺度图像构建、脉搏信号分层提取与融合的算法框架，并针对多尺度投影可以提高脉搏提取精度这一问题给出了理论分析与实测数据验证。在多尺度图像序列构建方面，提出不同宽高比的候选跟踪框生成方法、8邻域扰动重采样方法。在脉搏信号分层提取与融合方面，提出了基于类信噪比测度的信号筛选方法、基于高斯先验的凸组合融合方法、分层注意与融合机制等。在脉搏信号提取与融合网络结构研究方面，针对脉搏信号的特点，提出了基于空间-时间-通道联合注意机制、多视角卷积神经网络、可分离时空卷积和维度可分离注意机制、联合注意机制等。此外，本研究还针对现有目标跟踪算法对rPPG产生额外干扰的问题，提出了基于脸部地标点的rPPG信号友好的人脸跟踪方法。本项目为rPPG研究者们提供了一套新的理论与方法体系，为深入理解rPPG问题提供了更多的事实证据。

内容获取失败，请点击重试