图像序列中机动目标三维运动和结构的计算

综合视觉运动分析中的2类处理方法,选取图像中的角点作为特征点,在理论上证明了图像序列的光流场可以近似地用角点的位移场代替。利用已有文献中的建模思想,详细推导出递归计算机动目标三维运动和结构的非线性计算模型,采用广义卡尔曼滤波(EKF)递归地计算图像序列中机动目标的三维运动和结构。合成图像序列和真实图像序列实验结果表明该算法能取得较好的效果。     

面向视觉SLAM的快速均匀特征点提取方法CSCD

针对现有特征点提取方法中特征点提取结果冗余且均匀性差,导致SLAM定位精度及计算效率低的问题,提出了一种面向视觉SLAM的快速均匀特征点提取方法。在该算法的特征点提取部分,创新性地根据目标特征点提取数量对图像进行网格划分,使得特征点均匀分布在整幅图像中。同时,提出了一种兼顾全局和局部信息的均匀性评价标准,更符合视觉SLAM对特征点分布的要求。在NVIDIA Jetson TX2嵌入式平台上进行了实验验证,实验结果表明,新提出的特征点提取方法降低了特征信息冗余及特征信息缺失的概率,提升了SLAM定位精度,在计算资源有限的平台上具有实际应用价值。     

无人直升机视觉着陆中的运动状态估计算法

对无人直升机(UH)视觉着陆中基于视觉图像处理的运动状态估计问题进行了研究。介绍了视觉着陆原理,分析了运动估计、特征图像处理与着陆控制间的关系,推导并建立了UH相对着陆平台位姿估计算法、线速度与角速度估计算法。相邻两帧图像对应特征点像点位置为位姿估计算法提供数据,一帧图像特征点像点位置及其对应像点平移速度为线速度与角速度估计算法提供数据。利用UH着陆控制仿真数据模拟UH着陆运动过程中像点位置及其对应平移速度的视觉图像处理结果。仿真验证了运动状态估计算法,结果表明所提出的运动状态估计算法能有效地利用视觉图像处理结果数据估计出UH的位置、姿态、线速度和角速度。     

一种基于光流拟和的航拍视频图像全局运动估算方法

在多分辨率光流估算的基础上,提出了一种基于光流拟和的方法用于对航拍视频图像中的全局运动进行估算。在利用光流对全局运动模型拟和时,针对光流估算的不确定性以及图像中独立目标运动的干扰,提出了一种光流分块拟和与评估的方法用于剔除光流场中不利于全局运动估算的干扰数据,从而有效提高了全局运动的估算精度。航拍视频图像配准与独立运动检测的仿真结果均验证了算法的有效性。     

无监督深度学习单目视觉里程计研究

针对传统单目视觉里程计存在的尺度漂移和尺度不一致问题，提出了一种基于无监督深度学习的单目视觉里程计。首先，联合使用空间几何约束和图像相似性约束，得到长序列尺度一致的深度估计网络和鲁棒的光流估计网络；然后，对密集光流进行采样，得到精确的稀疏对应关系，减少尺度漂移；最后，根据改进的ORB-SLAM初始化方法，选择最优跟踪方式，结合深度信息进行尺度对齐，从而得到全局尺度一致的视觉里程计。在KITTI数据集上进行大量实验，结果表明，相较于ORB-SLAM2和基于深度学习的端到端的视觉里程计系统，该算法在通用评估指标方面性能有明显提升，验证了该算法的有效性。     

月面特征稀疏环境下的视觉惯性SLAM方法

月球车在执行科学探测任务过程中,其自身的高精度定位是一项亟需解决的关键问题。针对在特征稀疏的月面环境下的定位问题,提出一种视觉惯性融合的SLAM方法,将视觉测量与惯性传感器的信息利用位姿图优化方法融合,实现高精度的联合定位。针对特征稀疏环境下的前端视觉数据关联误差较大的问题,提出了一种基于四元树的光流跟踪算法,能够有效地跟踪鲁棒的特征点,提升了关键帧之间相对位姿估计的准确性。并且针对月面环境特有的恒星无穷远点干扰问题,提出一种高效的恒星点剔除算法,能够有效改善无穷远点导致的定位精度下降的问题。搭建了一套模拟月面环境的计算机仿真系统,并构建了多个月面环境视觉惯性SLAM仿真数据集,在不同的模拟月面场景下进行定位性能仿真验证,仿真测试结果表明本文算法的鲁棒性更强,具有更高的定位准确度。     

Use of land＇s cooperative object to estimate UAV＇s pose for autonomous landing

The research of unmanned aerial vehicles＇（UAVs＇）autonomy navigation and landing guidance with computer vision has important signifcance.However,because of the image blurring,the position of the cooperative points cannot be obtained accurately,and the pose estimation algorithms based on the feature points have low precision.In this research,the pose estimation algorithm of UAV is proposed based on feature lines of the cooperative object for autonomous landing.This method uses the actual shape of the cooperative-target on ground and the principle of vanishing line.Roll angle is calculated from the vanishing line.Yaw angle is calculated from the location of the target in the image.Finally,the remaining extrinsic parameters are calculated by the coordinates transformation.Experimental results show that the pose estimation algorithm based on line feature has a higher precision and is more reliable than the pose estimation algorithm based on points feature.Moreover,the error of the algorithm we proposed is small enough when the UAV is near to the landing strip,and it can meet the basic requirements of UAV＇s autonomous landing.     

一种改进的Lucas—Kanade光流估计方法

Lucas—Kanade算法是一种重要的光流估计技术,在计算机视觉领域有广泛的应用。但对于图像纹理欠丰富的区域,Lucas—Kanade算法光流估计结果较差。文章提出了一种改进的Lucas—Kanade算法：首先,计算跟踪置信度因子;然后,通过阈值检测剔除不可靠的光流估计结果;最后,进行邻域填充。该方法改善了纹理欠丰富区域的光流估计,对矮人模型图像序列进行了实验,证明了方法的有效性及优于Poelman方法和Chiba方法。     

A sequential detection approach to target tracking

Sequential hypothesis testing is investigated for multiframe detection and tracking of low-observable maneuvering point-source targets in a digital image sequence. The proposed multiple multistage hypothesis test tracking (MMHTT) algorithm extends tracks formed from sequentially detected target trajectory segments using a multiple hypothesis tracking strategy. The MMHTT algorithm does not require a probabilistic larger maneuver model. Computational efficiency is achieved by using a truncated sequential probability ratio test (SPRT) to prune a dense tree of candidate target trajectories and score the detected trajectory segments. An analytical performance evaluation is presented and confirmed by experimental results from an optical satellite tracking application     

The Estimation of the "Center of Gravity" of a Photon Density Profile in Noise

A model for an optical position estimation system is developed employing the photon Poisson process theory. The position estimate is based upon the definition of a center of gravity (CG) of the power density profile of the optical source on the focal plane. An estimator structure is derived using maximum likelihood estimates of the image profile. The resulting estimate of the CG is shown to be unbiased and its variance is obtained. The variance is shown to depend upon the signal energy and noise level as well as upon the distance of the center from the initial counting point. Thus, a composite estimation system is presented which reduces the variance and yet yields a simple structure. Studies on star estimation have yielded position accuracies better than 0.1 seconds of arc for a 2.5 visual magnitude star in a background of equivalent intensity.     

基于改进ORB-GMS-SPHP算法的快速图像拼接方法

针对传统图像拼接算法存在拼接速度慢、图像拼接有色差等问题，提出了一种基于ORB-GMS-SPHP算法的大视场多图像拼接方法。该方法首先利用高斯函数构建尺度空间，将高斯尺度空间划分为多个网格，在每个网格内借助FAST算法提取尺度空间特征点，使用BRIEF算法提取描述符并匹配，得到更加均匀分布的特征点；然后使用运动网格统计算法筛选匹配点；最后采用SPHP算法融合图像重叠区域，从而得到完整的拼接图像。将改进的ORB-GMS-SPHP算法与现有的传统特征点匹配算法在特征点匹配精度和特征点匹配速度进行对比与评价，验证了该方法特征点匹配速度快、精度高，并且可以保留更多的正确匹配点的特点。将该拼接方法与传统拼接方法在拼接速度、图像配准均方误差RMSE以及视觉主观判断拼接色差进行对比与评价，验证了该拼接方法具有较快的拼接速度、更高的拼接精度和无明显色差。该方法在2 736像素×3 648像素图像中,特征点匹配时间降低至6.463 s，图像配准精度RMSE降低至3.87。实验证明该方法特征点匹配速度快、精度高,且拼接精度高、无明显色差。     

基于改进Cycle-GAN的光流无监督估计方法CSCD

卷积神经网络为光流的计算提供了一种新的方式,但作为一种数据驱动技术,用于训练网络的大规模光流真值在现实世界中不易获取。为了解决这个弊端,基于Cycle-GAN的循环对抗机制,提出了一种光流无监督估计方法。首先,引入双判别器机制在生成器生成的光流样本的底层和高层特征上进行鉴别,迫使生成器提高光流生成的精度。其次,引入Spynet作为教师网络,在生成器训练前期对其进行指导,防止网络陷入模式崩塌。最后,改进损失函数,提出了光流一致性损失和轮廓一致性损失函数,进一步提升光流估计精度。实验结果表明,与现有的先进算法相比,提出的方法达到与有监督算法相同的精度水平。     

On-Line Vehicle Motion Estimation from Visual Terrain Information Part I: Recursive Image Registration

This paper addresses estimation of motion from the optical flow observed by an airborne down looking electro-optical sensor. The paper is in two parts. Part I addresses the development and analysis of a velocity-to-height ratio estimation algorithm and its principal error characteristics. In part II, it is shown how the information provided by the motion estimator can be integrated with additional on-board sensors to provide a complete autonomous navigation system. Part I as presented here is a summary version of the full length paper.? The algorithm implements recursive registration of successive images by using the gradient of a similarity function between them to control the tracking of their relative shift. The shift estimate provides velocity/height information. Substantial saving in memory and computation as compared to conventional full frame registration is achieved by using only a single line in the TV frame. Stochastic mathematical models for the image, terrain and vehicle velocity perturbations are used in the analysis. The choice of the most appropriate similarity function in the registration algorithm is addressed. Performance analysis indicates very small error variances, as illustrated by numerical examples.     

一种基于航摄图像特性的H.263压缩算法的改进

研究了航摄视频图像的特性,并根据航摄视频图像的特性,从预测初始搜索点、搜索策略和半像素精度运动估值方面改进了H.263标准中的运动估计算法。实验结果表明,针对航摄图像特性改进的算法比全搜索算法和三步搜索算法具有更高的搜索效率。     

基于人工神经网络的多模型目标跟踪算法

针对在目标跟踪中单模型跟踪算法难以应对目标运动形式的变化,而多模型跟踪算法存在结构固定、跟踪精度被非匹配模型削弱且模型切换缓慢的矛盾,文章提出了一种基于人工神经网络的多模型目标跟踪算法。通过分析目标几种基本运动模式的轨迹特点,归纳出目标运动轨迹的特征向量。利用训练好的BP神经网络对滑窗里的轨迹段进行运动模型识别,按结果进行跟踪模型切换,达到使跟踪算法实时适应目标运动状态的目的。仿真结果证明了该算法的有效性,且与传统的多模型算法相比,具有结构更加简单、更强的灵活性和拓展性的特点。     

基于PDR和磁场匹配的智能手机室内定位

目前，基于磁场的室内定位方法存在指纹采集无法快速建库的问题，导致匹配特征较少、无法快速进行匹配操作和航向角估计精度低。基于磁场梯度的快速近似，使用高斯牛顿迭代方法结合通过PDR测量的轨迹进行磁场轮廓匹配定位，提高了单点磁场指纹的可分辨性。基于步态模型的更新用作测量信息以改善C-INS的导航性能，提出了一种基于捷联惯性导航系统的行人航位推算算法。基于MFS参考位置，使用扩展Kalman滤波器控制PDR惯性导航的位置漂移误差，进一步提高了轨迹轮廓的精度。测试结果表明，该算法可以获得更好的位置估计和航向估计结果。     

Optical flow based guidance system design for semi-strapdown image homing guided missiles

This paper focuses mainly on semi-strapdown image homing guided (SSIHG) system design based on optical flow for a six-degree-of-freedom (6-DOF) axial-symmetric skid-to-turn mis-sile. Three optical flow algorithms suitable for large displacements are introduced and compared. The influence of different displacements on computational accuracy of the three algorithms is ana-lyzed statistically. The total optical flow of the SSIHG missile is obtained using the Scale Invariant Feature Transform (SIFT) algorithm, which is the best among the three for large displacements. After removing the rotational optical flow caused by rotation of the gimbal and missile body from the total optical flow, the remaining translational optical flow is smoothed via Kalman filtering. The circular navigation guidance (CNG) law with impact angle constraint is then obtained utilizing the smoothed translational optical flow and position of the target image. Simulations are carried out under both disturbed and undisturbed conditions, and results indicate the proposed guidance strat-egy for SSIHG missiles can result in a precise target hit with a desired impact angle without the need for the time-to-go parameter.     

动态背景下基于金字塔LK光流的运动目标检测

提出了一种动态背景下基于金字塔LK光流的运动目标检测方法。首先采用Shi-Tomasi角点检测法提取图像的强角点,再用金字塔LK光流法计算下一时刻这些特征点的对应位置,得到匹配的多组特征点对。利用这些特征点对建立仿射变换模型并计算仿射参数,对图像做仿射变换消除背景运动带来的误差,最后用帧间差分法处理得到前景目标。实验结果表明,本方法能够以较好的实时性、稳定地检测和提取动背景下的运动目标。     

Using natural features for vision based navigation of an indoor-VTOL MAV

The use of natural features for vision based navigation of an indoor Vertical-Take-Off-and-Landing (VTOL) Micro Aerial Vehicle (MAV) named Air-Quad is presented. Air-Quad is a small four-rotor helicopter developed at the ITE.Such a helicopter needs reliable attitude information. The measurements of the used MEMS gyroscopes and accelerometers are corrupted by strong noise. To be useful, the MEMS sensors have to be part of an integrated navigation system with aiding through complementary sensors like GPS or the computer vision module presented here.In the computer vision module, feature points are detected and tracked through the image sequence. The relative rotation and translation of the camera are estimated using the two-dimensional motion of the feature points.The three-dimensional points in the scene are modeled with the image coordinates of their first sighting and their inverse depths. Only these inverse depths are estimated for the feature points. An efficient sparse bundle adjustment algorithm is used to improve the estimation of the scene structure and the navigation solution.It is shown that the use of the computer vision module greatly improves the navigation solution compared to a solution based only on MEMS sensors.     

失效卫星远距离相对位姿估计与优化方法

针对远距离非合作慢旋目标位姿估计精度问题,提出一种融合图像超分辨与视觉SLAM的相对位姿估计方法。算法主要包含3个步骤：通过梯度引导生成式对抗超分辨技术,提升目标图像的质量以获取更多更高质量的特征点;构建特征数据库实现当前帧与特征数据库的匹配,提升旋转目标的特征跟踪稳定性;利用图优化对多帧图像进行联合位姿优化,消除累计误差,得到更为精确的估计结果。为稳定网络的训练,将自然进化算法引入到对抗训练中。为增强模型的泛化性和鲁棒性,实验中的数据集采用半物理仿真获得。实验结果表明,当等效距离为25 m且失效卫星以25（°）/s的速度旋转时,目标图像经超分辨网络增强后,能够实现连续稳定的长时间测量。