面向运动目标检测识别的高分辨光学压缩成像理论与技术

The emerging theory of compressed sensing has potentially powerful implications for the design of optical imaging devices and its application fields. In this project we aim to develop a high resolution compressive imaging system, which has the characteristics of small size, simple structure and low energy consumption. Based on this compressive imaging system, we will apply it into wide-area video surveillance fields and explore a completely new moving target perceptual modal. Actually, the state of art feature extraction method always need complex algorithm to reduce the data dimentation, which waste a lot of computation resources. In order to solve this problem, we propose to realize object detection and recognition algorithem with sparse imaging representation. Although compressed sensing theory suggests that one can recover a scene at a higher resolution than is dictated, this remarkable result comes with some important caveats, especially when practical issues associated with physical implementation are taken into account. In this project, we discusses compressed sensing in the context of optical imaging devices, emphasizing the practical hurdles related to building such devices, and offering optical algorithms for overcoming these hurdles. Coded aperture lens array are used to modulate the space light come from the scene. Low resolution detector arrays are joined together to obtain the sparse sampling image signals. Moving targets will be detect and recognize directly in compressive sensing image space. This new moving target perceptual modal, which based on compressive imaging data, not only can dramatically reduce sampling costs, energy consumption and alleviate communication and storage burdens with high resolution image performance. But also it can decrease the computational costs used for image signal processing and overcome the weakness of our vision algorithm which is light sensitive. Our research will change the moving target perceptual modal and make it more adaptive to the lightweight, network, and universal development direction.

课题突破现有光学成像模式，研制新型高分辨压缩成像系统，并基于该成像系统开展稀疏域运动目标感知新方法研究。解决传统成像系统"高速采样"再"压缩"所造成的采样资源浪费，以及现有视觉算法需要对高维图像数据进行降维处理的计算资源浪费问题。课题采用分块编码孔径透镜阵列实现大视场范围的目标非自适应线性投影，在小面阵、低分辨的图像传感器阵列上实现稀疏采样图像的高分辨成像。针对压缩成像系统的实际物理器件约束，开展测量矩阵非负补偿算法，光能非负条件下的图像快速重构算法研究，根据压缩成像系统的数据获取体制，探索稀疏域的运动目标检测识别算法，为压缩成像机理下的目标特征理解提供新思路。课题的研究成果能够有效缓解现有数字成像系统高速采样实现的压力，减少数据存储、传输代价，缩短信号后处理所需的时间和计算成本。为推进现有目标感知模式向轻量化、普适化和网络化方向的发展奠定坚实的基础。

课题在压缩感知理论的基础上，研制开发了一种结构简单、低成本、低能耗的新型压缩成像系统，并基于该系统展开稀疏域的运动目标感知新方法研究。该压缩成像系统能够解决传统成像先采样，后压缩所造成的采样资源浪费以及现有视觉算法需要对高维图像数据进行后信号处理的复杂计算消耗问题，具有重要的科学研究意义和实用价值。课题的主要研究内容和成果主要包括压缩成像系统的理论推导与系统原型的设计开发、压缩感知图像域的目标检测、识别与跟踪算法。首先，课题在理论推导的基础上，实现了投影矩阵的优化设计。搭建了一种基于数字微镜阵列(Digital Micromirror Device, DMD)分块控制和超分辨重建的并行可见光焦平面压缩成像系统。针对目前压缩感知直接成像技术存在采样时间长和计算复杂度高的问题，系统利用DMD 分区控制实现目标图像的分块采样，利用TVAL3算法进行图像重构，恢复出高分辨率原始图像。实验结果表明，DMD分块控制与超分辨重建相结合的方式可以明显降低压缩感知成像系统的计算量，避免了由于测量矩阵过大而带来的存储和计算问题，提高了采样速率，缩短了成像时间。在压缩成像系统的基础上，课题组直接在压缩感知图像空间进行运动目标的鲁棒检测与特定监视目标的识别。提出了压缩感知域的多高斯混合背景建模算法，实现了压缩域前景运动目标的实时检测。在检测到运动目标的基础上，通过L1范数最优化求解稀疏表达系数，并在粒子滤波框架下实现目标的实时跟踪。该目标跟踪算法能够在10%的采样率下，正确跟踪到目标物体。此外，课题还提出了一种压缩域单样本人脸识别算法。该算法是建立在随机投影和稀疏表示理论的基础上，针对单样本人脸识别问题提出的一种改进算法。课题组在通用的AR、Yale、MIT数据库上对该算法进行了实验验证。实验结果表明，课题提出的算法对环境光照变化、表情以及遮挡具有较强的鲁棒性，其识别率明显高于SRC以及ESRC算法。基于压缩成像技术的运动目标感知模式可以极大程度地降低图像信号的采样频率以及数据存储和传输代价，显著地缩短图像的信号处理时间和计算成本，为推进现有目标感知模式向轻量化、普适化和网络化方向的发展奠定坚实的基础。