基于热释电红外压缩采样实现的泛在人体智能监护理论与方法研究

Person surveillance has been applied for diverse circumstances. While conventional distributed optical sensing paradigms offer potential advantages to support many novel visual surveillance applications, there exist great challenges in fully integrating computer vision techniques with distributed wireless visual sensor networks. All of these challenges are due to the constrained computational, communication and energy resources in wireless sensor networks (WSNs), which prohibits the use of high-cost cameras with desirable resolutions as ubiquitous sensors deployed in a large-scale network. In this project, we intend to address these challenges via proposing a theoretically solid sensing paradigm, that is, compressive pyroelectric infrared (PIR) sensing. Inspired by the compressive sampling theory and the reference structure tomography (RST), we propose the compressive infrared sensing paradigm. In the context of the thermal radiation field, the human motion is sparse compared with the whole scene. Our model cast high-dimensional spatio-temporal information of the multi-level human motion as low-dimensional sensory data stream. In particular, sensing models are formulated in the form of spatial transformation mapping from object space to measurement space; the PIR sensors which response sensitively to the incident thermal flux provide the temporal information of the human motion. Based on this paradigm, we design the sensing models and systems for specific sensing tasks. The rationale of the compressive infrared sensing approach will be investigated by exploring the physically implemented compressive sampling with the visibility modulated PIR sensor arrays. Furthermore, to facilitate the physical implementation in the form of distributed sensor networks, we will research the issues of coverage scalability and sensing efficiency of compressive infrared sensing. In particular, the hierarchical compressive infrared sensing will be investigated for achieving scalable visibility coverage and fulfilling multi-target multi-level sensing requirements.

高效地融合运动行为线索，实现对人体运动行为的感知与响应，是人机交互、智能环境和泛在计算等领域共同关注的、富有挑战性的问题。本项目以压缩传感支配的参考结构层析成像理论为基础，以提高传感效率为研究主线，以实现泛在的人体监护为目标,通过参考结构实现红外辐射场的非同构映射，基于热释电红外传感器阵列对人体运动而引起的红外辐射场变化产生输出响应，将高维的人体运动时空信息映射为低维的传感器数据流，实现多层次人体运动信息的采样与压缩的一体化。在此基础上，建立限于特征的分层递阶人体运动感知模型，解决数据-目标关联问题，探讨基于多线索运动综合的协作感知，弱化多线索交叉和混叠产生的模糊性和不确定性，实现红外传感数据流的自动分割与运动推理。本项目解决了分布式无线传感器网络对计算资源与通信资源的约束问题，为大范围多目标多层次的人体运动智能监护提供重要参考。

本项目以实现泛在的人体智能监护为目标，以热释电红外压缩采样为感知手段，对压缩传感支配的几何参考结构层析成像进行了探索，分别从压缩传感模型理论和具体实现方法两个层面，重点研究了限于特征的压缩红外运动检测与跟踪技术，人体目标日常行为识别及异常行为监测技术，综合运用各种机器学习算法，突破了通信带宽以及计算资源受限等传感效率存在的瓶颈，形成了适用于分布式网络化实现方式的、具有多尺度信息获取能力的、支持多线索运动信息综合分析的新型传感模式。..此方面的研究，无论是对泛在的人体智能监护自身的突破发展，还是对压缩传感理论的应用拓展，都具有十分重要的学术意义，在智能环境、人机交互、智能监控及无线传感器网络等领域有广泛的应用价值。