【作者】 陈龙；

【导师】 郭宝龙；

【作者基本信息】 西安电子科技大学 ， 测试计量技术及仪器， 2013， 博士

【摘要】 辅助视觉导航技术是利用计算机视觉代替人眼对于周围景物环境进行感知，对感兴趣目标进行提取分割，对目标进行理解，并完成相应的视觉任务。目前，辅助视觉导航已经在盲人辅助视觉、视觉导航车辆和智能监控等领域得到广泛应用，在军事、航空航天等领域也得到了应用。但在实际应用中，仍存在一些问题。如在复杂环境中，存在大量冗余视觉信息时，如何快速准确地提取到用户感兴趣的目标，建立视觉处理通道，将运算集中在信息集中的感兴趣区域；在目标存在遮挡、旋转和出现相似目标时，现有的方法很难有效地完成跟踪和识别任务；在恶劣天气条件下，图像质量下降也同样会给后续的视觉任务增加困难。本文针对辅助视觉导航中的关键问题，对图像增强、目标检测、目标跟踪和目标识别等问题进行了深入研究，结合人眼视觉特性，实现了复杂环境中感兴趣目标提取，基于物理模型对雾霾、光照不均图像的快速增强，结合全局和局部特征在目标旋转、形变下进行有效识别，采用粒子滤波技术完成相似目标干扰、目标遮挡情况下跟踪等功能。最后，使用DaVinci系列DSP设计完成了一套空间非合作目标测量系统，具有体积小，可靠性高等优点，可以应用于导弹防御、海洋监控等其它领域。论文的主要工作及贡献归纳如下：1.给出了基于人眼视觉注意机制的目标检测和视觉信息处理通道模型。人类的视觉系统就是利用视觉注意机制，处理大量的视觉信息并及时做出反应。本文首先分析了人眼视觉特征，为视觉注意机制引入目标检测奠定了理论基础；提取视频图像中颜色、亮度和方向信息建立特征向量，并结合双目立体视觉的深度信息和光流场的运动向量，通过信息论的方式融合各通道特征图构建显著图。通过显著图可以更加有效地提取视频中的显著目标，建立视觉信息处理通道，通道内为感兴趣目标，通道外为背景信息，为后续的目标识别、目标跟踪奠定了基础。2.提出了一种基于双边滤波器的图像增强算法。恶劣天气下造成的图像质量下降，现有方法存在运算速度慢、光晕效应明显的问题。针对这一问题，对于有雾图像，根据大气物理散射模型，通过灰度腐蚀膨胀运算，去除了环境光估计中白色目标对场景恢复的影响；使用快速联合双边滤波算法，快速准确地计算大气幕V，并通过V获得透射率t，避免了普通双边滤波可能造成的光晕伪影。对于光照不均图像，将光照反射成像模型和物体的RGB色彩通道反射特性结合，提出了亮通道的概念，并从解析的角度给出光照分量的计算方式。通过快速联合双边滤波算法，准确得到了光照分量，最后通过光照反射成像模型求解目标的RGB通道反射系数。实验表明，本文算法恢复图像具有细节丰富、自然逼真、清晰度高等特点，相较其它增强方法，运算速度大幅提高，有利于算法的硬件实现。3.提出了一种结合局部和全局特征的目标识别方法。根据视觉生理学和心理学研究，人类视觉系统依赖于注意力选择机制，结合视觉记忆，能够快速的理解场景内容。Gist向量对于一幅场景图像可以快速地获取的全局意义上的场景要点语义，但对于具体目标缺少特征提取。视觉注意机制分割目标后，提取目标3D-SIFT特征点。局部特征针对图像中感兴趣区域，实现对目标细节的描述，并且使用3D-SIFT描述子满足目标旋转、缩放、光照变化等不变性。局部特征结合Gist全局信息，通过SVM支持向量机对特征向量进行训练。通过实验表明，该方法在目标识别和场景理解领域，均取得良好的识别效果。4.提出了一种基于模糊C均值聚类的粒子滤波方法。针对粒子滤波中出现多相似目标跟踪结果会发散和跟踪核函数窗宽固定的缺陷，以非线性、非高斯系统目标状态估计为主线，使用可变椭圆作为粒子区域，在粒子重要性重采样后，通过mean-shift算法获得每个目标的聚类中心，利用FCM聚类算法完成粒子聚类，获得相应目标的粒子子群，最后通过粒子子群估计各目标的最终状态并修正核窗口宽度。实验表明，与传统粒子滤波算法相比，该算法解决了传统粒子滤波的发散问题，对于目标发生旋转、尺度变化、遮挡等情况，能够准确地对多目标进行跟踪，减少了粒子数量，具有很好的实时性。5.综合视频编解码、网络通信、目标检测、立体视觉测量等技术，设计并实现了一套非合作目标位姿测量实验系统。该实验系统采用基于Hough变换的线特征提取与跟踪，实现对感兴趣区域准确定位；根据特征点坐标进行快速立体匹配与三维坐标计算，并建立目标坐标系，采用RANSAC算法提高相对位姿计算精度。系统硬件使用DaVinci系列DM6467T核心芯片，外围电路采用模块化设计。对非合作航天器自主交会过程模拟的半实物仿真，测量实验表明，相对位置测量精度小于±20mm，相对姿态测量精度小于±2°，测量速度能够达到10fps。系统主要特点：嵌入式设计，硬件实时检测跟踪，体积小、功耗低，可应用在航天器对接、海域、战场监视、视觉导航等领域。更多还原

【Abstract】 Visual navigation aid technology substitutes computer vision for human eyes toperceive the surrounding environment, then extracts and understands interestedtargets, and accomplishes the corresponding visual task. At present, visual navigationaid has already been widely used in visual navigation aid for the blind, visualnavigation for vehicles, intelligent monitoring and other fields. Moreover, it has alsobeen applied in military, aerospace and other fields.The key technology of visualnavigation aid in image treatment includes image enhancement, target detection,target tracking, target identification and so on. There are some outstandingalgorithms for each of these treatments, however, problems still exist. For example,when large visual information occurs in complicated environment, how to extractclient-interested targets accurately and quickly, establish visual treatment channel,and compute the interested area in the information set; when the targets are shaded,blocked, rotated and there are similar targets, the present methods are hard tocomplete the tracking and identification tasks; under terrible weather condition,image quality reduction will also provide difficulty for the subsequent visual tasks.This paper made deep study on the key points of visual navigation aid, such asimage enhancement, target detection, target tracking, and target identification. Itcombined human visual properties to extract the interested targets in the complicatedenvironment, and fast enhance the halo and non-uniformed illumination imagesbased on physical model. Then, overall and local characteristics were used toeffectively identify the targets even when they were rotated or deformed. Andparticle filter technology was adopted to complete the tracking task even when therewere similar targets or the targets were blocked. Finally, DSP design of DaVinciseries was used to form a set of spacial non-cooperative target measuring system,which could be applied in fields such as missile defence, and ocean monitoring dueto its small size and high reliability.The main achievements of this paper are as follows:1. It studied human visual attention mechanism. Human visual system usesvisual attention mechanism to handle huge visual information and make timelyresponse. This paper raised a method to establish target detection and visualinformation handling channel based on human visual attention mechanism. It firstlyanalyzed human visual characteristics, which laid theoretical basis for visual attention mechanism to introduce target detection; Secondly, it extracted color,luminance and direction information from the images to build feature vectors,combined them with binocular stereo visual motion vectors of depth information andoptical flow field, and established saliency map by fusing saliency maps of eachchannel through information theory. Then, salient targets could be more effectivelyextracted from the video through saliency map to build visual information treatmentchannel. Interested targets stayed inside the channel, while outside were backgroundinformation. This laid basis for the subsequent target detection and tracking.2. When comes to the quality-deducted images due to terrible whethercondition, the current methods are slow in computation and the results are withobvious halo. To solve the problems, this paper raised an image enhancementalgorithm based on bilateral filter. As for foggy images, and according toatmospheric physics scattering model, this paper eliminated the influence of whitetargets in the ambient light to the scene restoration through grayscale-erosion-dilationoperation; then used fast joint bilateral filter algorithm to quickly and accuratelywork out atmospheric veil V, gained transmissivity t through V, which avoided thehalo artifacts possibly resulted by ordinary bilateral filtering. As for non-uniformedillumination images, this paper combined light reflection imaging model with objectRGB color channel reflection characteristics to raise a bright channel concept, andgave the computation method of illumination components from an analytical point ofview. Accurate illumination components were gained through fast joint bilateral filteralgorithm. Finally, target RGB channel reflection coefficient was solved by lightreflection imaging model. Experiments proved that the images restored by thisalgorithm were full of details, natural, vivid, and clearly visible. Compared withother enhancement methods, the speed of this algorithm was accelerated, which wasbeneficial for the hardware implementation.3. According to visual physiological and psychological studies, human visualsystem relies on attention selection mechanism, which combines with visualmemories to fast comprehend scene contents. According to visual perceptionmechanism, this paper combined overall characteristics with local ones to identifythe targets and comprehend scene. Gist vector can quickly gain overall scene keypoint semantics from a scene image, but it cannot extract feature characteristics ofspecific targets in the scene image. Visual attention mechanism extracted target3D-SIFT feature points and combined them with Gist overall information aftersegmenting target, and trained feature vectors by supporting vector machine (SVM). Experiments proved that this method had gained excellent results in targetidentification and scene comprehension.4. When there are multi-similar-targets, tracking results of particle filter willdiverge and tracking kernel function window width is fixed. To solve the problems,this paper raised a particle filtering method based on vague C mean value clustering.It made non-linear, non-Gaussian system target state estimation as mainline, usedvariable ellipse as particle region, gained clustering center of each target throughmean-shift algorithm after particle importance resampling. Then, FCM clusteringalgorithm was used to complete particle clustering and get corresponding targetparticle subblock. At last, the finally state of each target was estimated throughparticle subblock and kernel window width was corrected. Experiments proved that,compared with traditional particle filter algorithm, this algorithm solved thedivergent problem existed in traditional particle filter, and reduced particle quantity.It can real-timely and accurately track multi-targets even when the targets are rotated,blocked or changed in scale.5. This paper synthesized technologies such as video coding and decoding,network communication, target detection, and stereo visual measuring, designed andrealized a set of non-cooperative target position measurement system. This systemextracted, tracked, and accurately positioned the interested region based on Houghtransform linear characteristics; then made fast stereo match and3D coordinatecalculation according to feature point coordinates, established target coordinatesystem, and enhanced the accuracy of relative pose calculation by using RANSACalgorithm. This system used DM6467T core chip of DaVinci series as hardware, usedmodular design in the peripheral circuit. Simulative semi-physical simulationexperiment of non-cooperative spacecraft autonomous rendezvous proved that therelative position measuring accuracy of this algorithm was within±20mm, relativeattitude measuring accuracy was within±2°, and the measuring speed reached10fps. The advantages of this system were: embedded design, hardware real-timedetection and tracking, small in size, low power consumption, it can be applied infields such as spacecraft docking, ocean and battle field monitoring, and visualnavigation.更多还原