面向智能养老视频监控的运动目标检测关键技术研究

Towards the massive application of intelligent video surveillance systems for elder care in the future, together with consideration of video data transmission, storage and privacy protection requirements, this project studies initialization of background models based on tensor completion with general applicability, constructs online optimized models using compressive sensing to reconstruct foreground objects with practicability, and proposes adaptive parameter setting approaches for the online optimized models with progressiveness. The project shall take advantage of manifold learning theory to find proper subspaces that have unique solutions for tensor decomposition, and present background initialization schemes based on tensor completion, which offers effective technical support for precise sensing of elder-care monitoring scenarios. Then, to meet the requirements of data volume in transmission and storage and detection accuracy of the foreground objects, the project shall establish the online optimized models for foreground reconstruction utilizing the initialized background model, and improve the online optimized models with constraints, for example, prior information in videos and characteristics like saliency of the foreground objects. Moreover, the project shall set adaptive compression ratio in the online optimized models for each frame of videos in virtue of prior knowledge, and enhance their adaptability to variability of foreground sparsity, which provides an effective solution to robust foreground detection under the compressive sensing framework. The study of this project shall provide theoretical basis and technological reserves for motivating the massive popularization and application of intelligent video surveillance systems for elder care in the multi-level elder service hierarchy, which shall be of penetrating significance to industrialization of elder service.

本项目面向未来智能养老视频监控系统大规模应用，考虑数据传输存储及隐私保密需求，从普适性出发研究基于张量填充的背景模型初始化，从实用性出发构建基于压缩感知的在线前景重建优化模型，注重先进性给出了基于自适应压缩感知的前景检测方案。利用流形学习理论找到张量分解具有唯一性的子空间，提出了基于张量填充的背景模型初始化方案，为准确感知养老监控场景提供技术支持；利用初始化背景模型给出基于压缩感知的在线前景重建优化模型，并结合先验信息和前景目标显著性等特性对优化模型进一步约束改进，可满足养老监控对传输存储数据量及对前景目标检测准确性的要求；提出借助先验信息自适应设置视频数据各帧的压缩比，增强对前景目标稀疏度随时间变化的适应性，为压缩感知条件下进行鲁棒的前景检测提供解决方案。本项目的研究成果可为推动智能视频监控技术在多层次养老服务体系中的大规模普及和应用提供理论依据和技术储备，对养老服务产业化具有深远意义。

随着全球人口老龄化不断加剧和人们医疗保健意识逐步提高，智能养老监控受到世界范围的广泛关注。基于视觉的智能养老监控由于其部署简单、信息丰富等优点在未来推广应用中必将具有广泛的普适性和可行性，然而，视频监控数据量巨大，对传输系统带宽提出较高要求，也给中心处理器处理、存储造成较大负担，且长期完整的视频数据存储也会带来用户隐私泄露的隐患。针对智能养老视频监控的应用推广需求，本项目瞄准智能监控在养老监控领域中亟需解决的关键问题，结合养老监控环境下运动目标的特殊性，兼顾数据传输存储及隐私保密需求，围绕着背景模型初始化、前景重建优化和基于压缩感知的前景检测开展了深入的研究工作。为准确估计背景模型，利用流形优化理论，提出了基于流形优化的张量填充背景模型初始化方案，为准确感知养老监控场景提供技术支持；基于低秩子空间学习理论，提出了自适应在线背景模型初始化方案，可获得适应场景动态变化的背景模型估计，为准确感知场景状态提供理论依据；考虑到数据传输存储及隐私保密需求，提出了基于压缩感知的在线前景目标重建优化方案，结合前景目标先验信息和显著性，以在线方式从压缩数据中重建前景目标，可满足养老监控对传输存储数据量及前景目标检测准确性的要求；利用张量分解理论，提出了基于张量分解的在线前景重建方案，有效提高了前景重建对环境变化的适应性，为及时准确掌握监控目标信息提供技术支持；考虑到前景目标稀疏度的时变性，提出了基于自适应压缩比的前景检测方案，利用先验信息动态调整压缩比，在减小采集数据量的同时保证前景目标准确重构，以增强前景检测对视频图像的适应性；在压缩感知域引入神经网络，提出了基于改进脉冲耦合神经网络的压缩域前景检测方案，为压缩感知条件下进行鲁棒的前景检测提供解决方案。本项目研究成果不仅可为智能视频监控产业化发展奠定理论基础，而且对加快我国智慧城市和智慧社区建设、促进我国国民经济和社会和谐发展具有重大意义。