基于压缩感知的WMSN编码传输与视频重构技术研究

There are such challenges as low complexity coding, reliable transmission and efficient reconstruction of video information in wireless multimedia sensor networks (WMSN). Therefore, the proposal investigates the multi-hop relay transmission feature of WMSN and the compressible and layered features of multiple description coding (MDC) video in deep, and then proposes one multi-path coded transmission optimization model based on distributed compressed sensing for WMSN, which is on the basis of multiple description coding based multi-path routing, distributed compressive sensing and multi-view video reconstruction technologies. The optimization objective of proposed optimal model is to reduce video reconstruction distortion rate with constraints such as node resources, link bandwidth, packet loss, transmission delay, layered video compression and others. The proposal focuses on a serial of problems such as distributed compressed sensing joint multiple description coding, inter-frame similarity measure between multi-nodes, video reconstruction at sink node with adaptive weighted model etc. Moreover, the swarm intelligence algorithm is introduced to discuss the proposed multi-path coded transmission and video reconstruction model of MDC video. Furthermore, we use network simulation software and actual experimental platform to verify various mechanisms proposed in this project. Based on the global optimization and local control, the project proposes a distributed sub-optimal algorithm. Finally, the project concludes a series of optimal scheduling strategies with practical value and provides a new solution for the researches and applications of coded video transmission and effective video reconstruction technology in WMSN.

针对WMSN中视频数据面临着低复杂度编码、可靠传输与高效重构等技术挑战，本项目结合多描述视频(MDC)多径编码传输、分布式压缩感知和异构视角多视频融合技术，对视频数据在WMSN中多路径协作编码传输与高效重构进行理论研究。项目以提高视频重构质量为优化目标，以节点资源、链路带宽、链路丢包、传输延时、视频分层、视频帧相似度为约束，重点围绕多径冗余路由MDC编码传输策略、多节点异构视角视频帧相似性度量和压缩感知联合MDC视频重构等问题，构建基于分布式压缩感知的MDC编码传输与视频重构的优化理论模型。在全局优化和局部控制的基础上，提出符合工况的实用算法：包括压缩感知与多描述联合编码算法、基于自适应加权模型的解码端重构算法、基于相似性度量的多节点异构视角视频融合机制等，通过仿真软件和测试平台对各项指标进行量化分析以验证算法性能，为WMSN中数据编码传输与视频高效重构机制的研究提供一种全新解决思路。

本项目深入分析WMSN中多节点异构视角数据相关性、低分辨节点图像可重构性、固定无线多跳和移动Sink采集特征，提出了基于多视角融合技术的无线多媒体传感器网络系统架构，构建多视角无线传感网场景下的人体重识别（Re-ID）方法框架，并验证了平行阵列视角和环形阵列视角下的视频融合方法。针对无线多媒体传感器节点采集图像的低分辨率特性，本项目提出了两种新型的图像超分辨率重构算法以提高WMSN节点采集的源图像质量，分别应用循环生成对抗网络（CycleGAN）和双重残差注意力对偶回归网络实现高效率的接收图像重构。针对无线多媒体传感器网络的数据采集效率问题，克服传统无线多跳数据传输方式固有的Hotspot问题，本项目研究了无线多媒体传感网的新型移动数据采集与优化数据转发方法：将多移动Sink节点协同数据采集框架建模为具有NP-Hard难度的线性规划问题，采用自组织映射网络（SOM）无监督学习方法进行迭代求解，最终实现基于Dubins曲线的平滑路径规划与基于流函数的移动避障设计。本项目将相关研究成果应用于人体行为姿态识别领域，针对异常跌倒、坐姿检测等实际需求，开发了相应的在线监测算法和软件系统。本项目共发表学术论文15篇，其中SCI/EI收录学术论文8篇，授权国家发明专利2项，联合培养硕士研究生2名。本项目的研究成果为多视角WMSN的移动数据采集传输与图像超分辨率重建提供了一些新思路新方法，为部署和应用WMSN提供了技术支撑。