【作者】 张文豪；

【导师】 徐明亮；

【作者基本信息】 郑州大学 ， 计算机科学与技术， 2025， 硕士

【摘要】 非视域(Non-Line-of-Sight,NLOS)成像是一种通过分析光的传播特性,对不可直接观察的目标进行重建的成像技术。其核心原理是利用环境中散射或反射的光线,从遮挡物表面或其他介质中提取隐藏场景的信息。这项技术在安全监控、自动驾驶、医疗成像、无人系统等领域具有广泛的应用前景。在安全监控中,非视域成像可用于探测建筑拐角后的隐蔽威胁;在自动驾驶中,该技术可提前识别遮挡物后方的潜在危险物体;在医疗领域,此技术可辅助实现非侵入式诊断。由于非视域成像技术所采集中继面信号包含的可用信息较少,当前的研究主要聚焦于挖掘额外信息以提升成像质量。现有方法包括利用可编码光源、高时间分辨率探测器、场景先验知识,以及光的波长、偏振和相干特性,并结合深度学习技术进一步优化。然而,该技术仍然受到单一模态的局限,未能充分挖掘多种模态的互补优势。此外,在实际应用中,仍缺少用于实时重建的软硬件一体化的解决方案。本论文围绕跨波段偏振非视域成像技术展开研究,关注更多模态信息,以提升成像的鲁棒性和精度。主要工作包括以下几个方面:(1)制作了跨波段偏振非视域数据集,搭建实验平台,设计数据采集方案,利用可见光偏振成像设备与长波红外偏振成像设备采集数据,创建跨波段偏振数据集,为后续算法研究提供基础。(2)设计了偏振增强的长波红外非视域成像方法,提出了一种使用偏振模态进行非视域重建的逐级信息补充方法。在红外强度图的基础上,逐步引入偏振模态数据,并结合大核卷积与混合并行注意力机制确保附加信息可以得到有效利用,优化了偏振特征提取与使用,重建图像显示出增强的纹理细节,提高了重建隐藏目标的精度。(3)设计了跨波段图像转换的非视域成像方法,基于生成对抗网络模型,采用端到端的多任务学习策略,实现跨波段图像转换与增强,在保留长波红外波段高镜面反射优势的情况下,利用红外着色技术,建立了隐藏目标与其自身颜色之间的映射关系,补偿了红外图像中颜色信息的缺失。(4)开发了动态环境下的实时非视域成像系统,使用车载双目相机、无人狗、虚拟现实眼镜和其他网络终端设备组成非视域实时成像系统,部署算法模型推理重建,为动态场景中的非视域成像提供了解决方案。更多还原

【Abstract】 Non-Line-of-Sight(NLOS)imaging is an advanced imaging technique that reconstructs scenes hidden from direct view by analyzing the propagation characteristics of light.Its fundamental principle lies in extracting information about occluded scenes from scattered or reflected light off surfaces or other media in the environment.This technology holds significant promise across various domains,including security surveillance,autonomous driving,medical imaging,and unmanned systems.In security surveillance,NLOS imaging can detect concealed threats around building corners;in autonomous driving,it enables early identification of hidden obstacles;and in the medical field,it facilitates non-invasive diagnostics.Due to the inherently limited information contained in the relay surface signals captured in NLOS imaging,current research primarily focuses on exploiting additional modalities to enhance reconstruction quality.Existing approaches incorporate techniques such as structured light sources,high temporal resolution detectors,prior knowledge of scenes,and optical properties including wavelength,polarization,and coherence,often combined with deep learning for further optimization.However,these techniques still face limitations due to their reliance on single modalities and fail to fully leverage the complementary advantages of multi-modal information.Moreover,practical applications still lack integrated software-hardware solutions capable of real-time reconstruction.This thesis investigates cross-spectral polarized NLOS imaging with a focus on multi-modal information integration to improve imaging robustness and accuracy.The main contributions are as follows:(1)A cross-band polarization NLOS dataset was created by constructing an experimental platform and designing a data acquisition scheme.Data were collected using visible-light polarization imaging devices and long-wave infrared polarization imaging devices,forming a cross-band polarization dataset that serves as a foundation for subsequent algorithm research.(2)A polarization-enhanced LWIR NLOS imaging method was developed,introducing a progressive information integration strategy that incorporates polarization modalities into the reconstruction process.Building upon infrared intensity images,polarization data were gradually integrated,supported by the use of large-kernel convolutions and a hybrid parallel attention mechanism to effectively exploit the supplementary information.This approach improves the extraction and utilization of polarization features,leading to enhanced texture representation in the reconstructed images and increased accuracy in recovering hidden targets.(3)A cross-band image translation-based NLOS imaging method was designed using a generative adversarial network model.By adopting an end-to-end multi-task learning strategy,the method achieved cross-band image conversion and enhancement.While preserving the high specular reflection advantage of the LWIR band,infrared colorization techniques were employed to establish a mapping between hidden targets and their inherent colors,compensating for the absence of color information in infrared images.(4)A real-time NLOS imaging system for dynamic environments was developed.This system integrates a vehicle-mounted stereo camera,an unmanned quadruped robot,virtual reality headsets,and other networked devices to enable real-time NLOS imaging.The proposed algorithm was deployed for inference and reconstruction,providing a reliable solution for NLOS imaging in dynamic scenarios.更多还原