单对立色流运动下的抠图跟踪

当目标与周围背景相似性较高,且目标运动形式复杂时,很难精确跟踪目标。针对传统目标跟踪方法因目标外观的不精确建模致使模型退化而产生漂移的问题,提出了一种单对立色流特征下的抠图跟踪方法。首先,根据彩色视频帧包含的丰富颜色信息,将单对立色颜色编码方式与Lo G兴趣点检测器结合,获取颜色兴趣点作为目标表示集合;其次,在目标运动预测阶段,通过估计相邻帧兴趣点的流匹配关系,获得预测帧的前景目标兴趣点估计;最后,利用获得的前景兴趣点估计进行抠图,重新划分当前帧的前背景标记,并对模型进行更新。该算法在Segtrack视频跟踪数据集上进行验证,定性定量分析了跟踪目标快速运动,目标形变和光照干扰下的跟踪效果。实验结果表明该算法可有效提高形变目标跟踪的准确性,优于当前目标跟踪的先进算法。

基于MAP-EM的多帧遥感影像超分辨率技术研究

文中把MAP联合求解模型应用到遥感图像序列的超分辨率重建中来,并针对其在边缘保持能力和收敛复杂度方面的局限性,做出如下的改进:运动估计阶段使用SIFT特征算子和光流法相结合提取运动矢量;引入一个二进制掩膜变量,减少模糊边缘效应;迭代重构采用EM算法,使最终结果收敛到一个稳定的最优解。最后的实验结果也表明,针对具有挑战性的低分辨率遥感图像序列的输入,我们的方法可以产生更加清晰和更高分辨率的重建结果。

Vision-Based Obstacle Avoidance in Sidewalk Environment Using Top-View Transform and Optical-Flow

This paper presents a vision-based obstacle avoidance method for blind pedestrians in an outdoor sidewalk environ- ment. Unlike many existing travel-aid systems using stereo-vi- sion based methods, the proposed method is able to get obstacle position as well as user motion information by just one monocu- lar camera fixed at the belly of the user. To achieve this goal, a top-view transformation of the road image is used for obstacle detection and user motion estimation, based on which a grid map is generated for navigation. For detection part, the bottom points of erect obstacles are detected by extracting local-maxima and minima on the top-view image while user motion is estimat- ed by analysing the optical flow vector field in the user sur- rounding area. For the obstacle avoidance part, a step score is calculated on the grid map for evaluating the safety of next moving step. Experiments with several sidewalk video-clips show that the proposed obstacle avoidance method is able to provide useful guidance instructions under certain sidewalk environments.

Estimation of the Hydrodynamic Regime of the Water Movement under the Influence of the Atmospheric Processes in the Bering Sea and the Sea of Okhotsk

The given investigation presents the results of estimating the water circulation in the water area of the Bering Sea and the Sea of Okhotsks, considering the influence of various types of the atmospheric processes. To solve the given problem it is used a hydrodynamic model calculating the integral functions of the flow from the surface to the bottom. By results of calculations, the maps of the integral water circulation were built for the following types of atmospheric circulation： ＂north-western＂ and ＂okhotsk-aleutian＂. In accordance with the performed calculations for the water area being studied, the hydrodynamic structures are distinguished both non-depending and depending on the type of the atmospheric circulation. The non-depending structures are characterized by the cyclonic activity in the Bering Sea and the Sea of Okhotsk in whole. Hydrodynamic structures depending on types of the atmospheric circulation have their peculiarities in the spatial-temporal distribution.

基于点线特征的单目同步定位与建图初始化方法

针对目前单目同步定位与建图(SLAM)初始化方法鲁棒性差、精度低的问题,提出了一种基于点线特征的单目SLAM初始化方法。首先,在两帧图像之间提取线特征,并进行特征匹配。然后,通过最大化投影线特征的重叠长度优化两帧图像之间的初始旋转和平移矩阵。最后,用滑动窗口增加初始化的图像帧数,利用多帧图像的信息和约束关系以及全局光束平差法优化初始地图和估计的相机关键帧位姿。在TUM和OpenLORIS数据集上的测试结果表明,相比传统初始化方法,本方法的鲁棒性更好、精度更高,能在具有挑战性的场景中快速完成高精度初始化。

室外环境下基于点线特征的视觉-惯性里程计和全球卫星导航系统定位算法

为了解决视觉-惯性里程计(VIO)在室外复杂环境中图像特征信息不足和累计误差的问题,设计提出了一种基于点线特征的单目VIO和全球卫星导航系统(VIO-GNSS)定位算法。首先,提出了一种基于几何约束的最小化策略对提取的线特征进行前后帧间约束匹配,改进了基于点线特征的单目VIO算法;在此基础上,建立了改进的VIO-GNSS松耦合定位模型,利用GNSS全局观测值对VIO位姿估计的累计误差进行校正和抑制。在KIITI数据集和实测数据集中进行了实验验证,结果表明,与几种同类算法相比,提出算法在GNSS信号良好/间断/拒止的大型室外复杂环境中均可保持良好的精度和鲁棒性,且满足驾驶定位的实时性要求。

面向动态场景的语义视觉里程计

针对传统视觉同时定位与地图构建(vSLAM)相机跟踪模块在动态环境中无法精确定位的问题,提出一种基于语义的视觉里程计。首先,在利用金字塔Lucas-Kanade光流追踪匹配帧间特征点的同时,对图像进行像素级的语义分割。然后,将语义信息与几何特征紧密结合用以准确地剔除图像中的外点,使得位姿估计和建图仅依靠图像中值得信赖的静态特征点。最后,提出了一种多尺度的随机抽样一致(RANSAC)方案,对匹配点进行步进采样,每步使用不同的尺度因子,在降低外点检测时间的同时,提高了外点检测的鲁棒性。在TUM数据集上的实验结果表明,在高动态序列中,相比于ORB-SLAM2,本文方案的绝对轨迹误差和相对位姿误差改善了90%以上,而相比于同类型的DS-SLAM,本文方案在降低外点检测时间30%~40%的情况下,提升了位姿估计的鲁棒性。