智能驾驶前视全向红外立体视觉理论与自然感彩色化方法研究

Due to the advantages of all-weather work, strong anti-glare, good dehazing, the automotive intelligent driving system (AIRIDS) has become a hotspot of automotive situation awareness for intelligent driving recently. However, the vision adaptation and scene comprehension are less considered in current research, and there are still many problems existing, including large forward-looking blind area, no depth information and low-level scene identification. Aiming at the development requirements of intelligent driving technology, the fundamental research is carried out about the theory of forward-looking omnidirectional infrared stereo vision and the method of natural-appearance colorization. Firstly, introducing non-similar imaging theory into the stereo vision, the mechanism of forward-looking omnidirectional infrared stereo vision is explored which can permit a complete hemispherical field of view to be imaged. Secondly, adopting the theories of coordinate transformation, polar curve geometry, etc, the calibration means and stereo matching approach are discussed which suits for the forward-looking omnidirectional infrared stereo vision. Thirdly, utilizing the techniques of supervised learning and depth color modulation, the natural-appearance colorization method with depth perception is studied. This research contributes to solving the problems of acquiring large-airspace scene, abstracting depth information and percepting colorization image for intelligent driving, and will help to expand the application fields of forward-looking omnidirectional infrared stereo vision, such as intelligent transportation, automatic driving, augmented reality, military reconnaissance, and so on.

车载红外智能驾驶系统（AIRIDS）因具有全天候工作、抗眩光能力强、穿雾透霾本领大等优势，而成为目前智能驾驶行车态势感知研究的热点。然而，当前AIRADS研究较少考虑视觉适应性和对场景的理解能力，存在前视盲区大、深度获取难、图像辨识度低等问题。为此，着眼智能驾驶技术发展需求，本项目拟开展前视全向红外立体视觉理论与自然感彩色化方法基础研究。①将非相似红外成像理论引入到立体视觉中，探索可获取半球空域场景深度的前视全向红外立体视觉新机制；②运用坐标变换和极曲线约束等理论，探究前视全向红外立体视觉标定和匹配新手段；③采用监督学习和深度色彩调制技术，研究兼具空间立体感和色彩自然感的红外图像彩色化新方法。研究成果有助于解决智能驾驶中的大空域场景获取、深度信息提取和图像彩色化感知等问题，也对拓展前视全向红外立体视觉在智能交通、自动驾驶、增强现实、军事侦察等军民领域的应用具有重要意义。

车载红外智能驾驶系统（Automotive Infrared Intelligent Driving System, AIRIDS）因具有全天候工作、抗眩光能力强、穿雾透霾本领大等优势，而成为智能驾驶中行车环境前视态势感知研究的热点。然而，当前AIRIDS研究较少考虑如何增大感知空域、获取前视场景深度信息、提高行车场景辨识度等问题。为此，本项目将非相似成像理论引入到红外立体视觉技术研究中，开展了前视全向红外立体视觉理论与自然感彩色化方法研究，丰富和完善了立体视觉研究体系。.（1）构建了可描述80°~180°视场的非相似红外立体视觉模型。基于等距投影成像思想论证并推导了非相似红外成像参数约束关系，构建并分析了红外立体视觉数学模型和视差模型；将立体视觉研究的响应波段从可见光波段（0.38μm~0.78μm）拓展至长波红外波段（8μm~14μm），将立体视觉系统的视场角从普通视场（一般≯40°）拓展至前视全向视场（180°）。.（2）提出了前视全向红外立体视觉高精度标定方法。建立了非相似红外立体视觉标定模型，提出了基于三步法参数优化的标定思路；设计并研制了棋盘式自适应双波段标定板，提出了基于亚像素边缘非线性拟合的角点检测方法；完成了180°视场红外立体视觉系统的标定，精度优于0.5个像素。.（3）提出了前视全向红外立体视觉精确匹配方法。通过推导超大视场双目极线约束关系，分析了双目图像匹配窗口形变特性，确定了超大视场红外立体视觉双目图像同名像点匹配约束关系；提出了基于可靠性检测的同名像点亚像素匹配方法，并实验验证了其有效性。.（4）提出了前视全向红外体视场景自然感彩色化方法。建立了“微光图像先色彩传递，再与红外图像融合并补充上色”的自然感彩色化策略，提出了基于超像素灰度特征和色彩聚类联合约束的超像素匹配机制，完成了微光图像的自适应色彩传递；采用基于RPCA的红外图像自然感彩色化方法，提高了彩色化后图像的双波段特征；建立了相似场景和重点目标红外独立信息的上色模型，实验验证了算法有效性。.最后，完成了原理系统的研制，并通过开展室内外实验，进行了系统的实验验证。