基于嵌入式GPU的汗孔识别算法并行设计

基于NVIDIA Jetson Tx2平台,结合OpenCV计算机视觉库与计算统一设备架构(CUDA)程序设计,对汗孔特征提取与匹配算法实现了并行设计.实验结果表明:并行设计算法能够实现最多180倍的加速,推动指纹匹配算法在嵌入式系统领域的应用.     

基于FPGA的稠密光流计算系统

高质量的稠密光流算法计算复杂度很高,因此计算速度成为制约其在实际系统中应用的重要原因。针对这一问题,利用现场可编程门阵列（FPGA）的细粒度并行特性,实现了一种高质量的稠密光流算法CBG（Combined- Brightness-Gradient）的硬件加速器。实验结果表明,在FPGA工作频率200 MHz,计算全部像素对应的光流信息的情况下,该系统处理分辨率为316×252的图像序列的帧频可达40 frame/s。     

基于CUDA的位并行近似串匹配算法

为满足文本检索、计算生物学等领域海量数据匹配对高性能计算的要求,提出一种基于计算统一设备架构(CUDA)的位并行近似串匹配算法。结合图形处理器(GPU)的高并行计算结构及存储带宽特性,通过优化数据存储方式,实现并行化动态规划矩阵算法(BPM)的加速,并对加速性能进行对比测试。实验结果表明,BPM算法通过GPU加速能获得20倍左右的加速比。     

一种基于光流和能量的图像匹配算法

结合光流与图像信息,提出一种获取稠密视差的图像匹配算法.首先对于基线较大的左右图像,在多分辨率框架下采用由粗到精的策略计算光流,从而实现大偏移量时的光流获取.其次为了避免光流在图像边界上的不可靠性,通过光流计算所得的光流场作为初始视差图,采用基于能量的方法依据对应的图像梯度场对光流场内部进行平滑并保持边缘的不连续性,最终得到精准稠密的视差图.实验验证,该方法是一种行之有效的图像匹配算法.     

基于DEM的Priority-flood并行填洼算法优化

针对大尺度数字高程模型无法适应单机内存，导致单机串行填洼算法无法计算的情况，对Barnes提出的并行PF填洼算法加以优化。基于spark实现了由Barnes提出的并行PF填洼算法，同时针对单张DEM未切分的情况，对该算法加以改进，设计了带光环的切分策略等一系列方法，将原算法一二阶段的同步处理变为异步，节约了原算法耗时。在进行2 600亿单元(10 m数据集)的填洼实验中，该方法与原方法填洼结果一致，且比原算法缩短了37%的处理时间，提高了并行填洼的计算效率。     

基于CLG光流算法的云的运动分析与研究

将光流算法应用到云的运动分析中,同时探讨光流变化与云的运动之间的关系.文中采用了局部与全局(CLG)相结合的光流算法分析云的运动.CLG算法同时其备局部光流算法和全局光流算法的优点,利用 CLG光流算法能得到鲁棒而且稠密的云的运动流场.文中首先详细分析了CLG 三种光流算法:空间CLG、时空CLG、非线性CLG光流算法.然后将这三种算法应用到云的运动视频中,并对三种CLG光流算法红云的运动上进行了分析和比较.实验证明,光流算法对于测量云的运动有良好的效果,云的运动与光流之间具有正比关系.     

SMT-PAAG下的Harris角点检测与匹配算法

结合多核处理器SMT_PAAG的平台特性,实现基于数据并行和任务并行的Harris角点检测与匹配算法。在SMT-PAAG仿真器上对其算法进行验证,根据加速比和效率两个性能指标对实验结果进行分析,结果表明SMTPAGG上Harris角点检测与匹配算法的并行化实现效果显著。     

一种用于智能车环境探测的立体视觉传感器

设计了应用于户外智能车环境感知的立体视觉系统,并提出了两种立体匹配算法,分别实现不同的功能.基于多特征提取的稀疏匹配算法首先通过图像局部增强处理减弱光照和场景变化的影响,然后在对极线引导和连续性约束下提取立体图像的角点特征和边缘特征进行匹配,能够实时探测环境的三维概貌,并突出环境中的障碍物信息,实现辅助车辆自主导航的功能;基于特征引导的稠密匹配算法采用了最小割/最大流算法,并利用多特征匹配的结果来剔除稠密匹配中的成块误匹配,能够重建出未知环境详细的三维信息,实现三维可视化功能.最后通过试验验证了两种算法的有效性.     

线云隐私攻击算法的并行加速研究

线云定位方法能保护场景隐私,但也存在被隐私攻击算法破解的风险。该攻击算法能从线云恢复近似点云,但其计算效率较低。针对该问题,提出了一种并行优化算法,并对其运行时间和加速比进行了分析。具体来说,分别采用SPMD模式和流水线模式实现了CPU多核并行和GPGPU并行。然后,进一步结合数据并行模式实现了异构计算,以达到最高的并行度。实验结果表明,并行优化算法加速比最大为15.11,最小为8.20;相比原算法,并行优化算法的还原点云相对误差控制在原误差的0.4%以内,保证了算法的精度。该研究对线云隐私攻击算法以及其他密度估计问题、不同场景下的线云隐私保护算法等有重要意义和参考价值。     

基于深度匹配的由稀疏到稠密大位移运动光流估计

针对非刚性大位移运动场景的光流计算准确性与鲁棒性问题, 提出一种基于深度匹配的由稀疏到稠密大位移运动光流估计方法. 首先利用深度匹配模型计算图像序列相邻帧的初始稀疏运动场; 其次采用网格化邻域支持优化模型筛选具有较高置信度的图像网格和匹配像素点, 获得鲁棒的稀疏运动场; 然后对稀疏运动场进行边缘保护稠密插值, 并设计全局能量泛函优化求解稠密光流场. 最后分别利用MPI-Sintel和KITTI数据库提供的测试图像集对本文方法和Classic + NL, DeepFlow, EpicFlow以及FlowNetS等变分模型、匹配策略和深度学习光流计算方法进行综合对比与分析, 实验结果表明本文方法相对于其他方法具有更高的光流计算精度, 尤其在非刚性大位移和运动遮挡区域具有更好的鲁棒性与可靠性.     

医学图像可视化自动编程平台的设计

为了充分利用现有的各种医学图像处理算法,避免重复开发,提高开发效率,设计并实现了一个基于OpenCV图像处理基础算法库的医学图像可视化自动编程平台。该平台在开放源代码的基础上,充分整合了OpenCV算法库,避免了算法的重复开发。通过对可视化编程技术的研究,实现了该平台中编程过程的可视化与自动化。同时利用OpenCV平台无关的特性,使用makefile文件来控制生成程序代码的编译,实现了生成的程序代码的可移植性。并讨论了平台的整体结构设计与各个模块具体实现。最后给出一个开发实例,证明了该平台在算法测试和开发方面的高效性与简洁性。     

基于稠密光流和边缘特征的烟雾检测算法

为了克服传统火灾烟雾检测技术的缺陷,提高视频烟雾检测算法的检测率,通过观察烟雾运动的特性,提出一种基于稠密光流和边缘特征的烟雾检测算法。该算法首先利用混合高斯背景建模和帧差相结合的方法提取运动区域,然后将此运动区域池化为上、中、下三部分,并在每个池化区域提取光流矢量特征和边缘方向直方图。考虑到烟雾运动在时域中的连续相关性,提取相邻三帧的烟雾特征向量以提高算法的鲁棒性。最后使用支持向量机进行训练和烟雾检测。实验结果表明,该算法在测试视频集上准确率超过94%,与现有方法相比,能更好地适应实际应用中复杂的环境条件。     

An improved algorithm of median flow for visual object tracking and its implementation on ARM platform

An improved algorithm of median flow used for visual object tracking is described. The improvement consists in adaptive selection of aperture window size and number of pyramid levels at optical flow estimation. It can increase the tracking efficiency as compared to the basic algorithm, especially when dealing with small and low-contrast objects. The proposed version of the algorithm has been implemented using OpenCV library and tested on OMAP 35x EVM and BeagleBoard-xM based on Texas Instruments OMAP3530 and DM3730 processors, respectively. Analysis of improved median flow was performed over actual video sequences. The results obtained show versatility and computational robustness of the algorithm, which makes it promising for embedded application based on ARM processors.     

Estimating the focus of expansion in a video sequence using the trajectories of interest points

In this paper, we present a new algorithm for the computation of the focus of expansion in a video sequence. Although several algorithms have been proposed in the literature for its computation, almost all of them are based on the optical flow vectors between a pair of consecutive frames, so being very sensitive to noise, optical flow errors and camera vibrations. Our algorithm is based on the computation of the vanishing point of point trajectories, thus integrating information for more than two consecutive frames. It can improve performance in the presence of erroneous correspondences and occlusions in the field of view of the camera.The algorithm has been tested with virtual sequences generated with Blender, as well as some real sequences from both, the public KITTI benchmark, and a number of challenging video sequences also proposed in this paper. For comparison purposes, some algorithms from the literature have also been implemented. The results show that the algorithm has proven to be very robust, outperforming the compared algorithms, specially in outdoor scenes, where the lack of texture can make optical flow algorithms yield inaccurate results. Timing evaluation proves that the proposed algorithm can reach up to 15fps, showing its suitability for real-time applications.     

Differential optical flow applied to automatic facial expression recognition

This work compares systematically two optical flow-based facial expression recognition methods. The first one is featural and selects a reduced set of highly discriminant facial points while the second one is holistic and uses much more points that are uniformly distributed on the central face region. Both approaches are referred as feature point tracking and holistic face dense flow tracking, respectively. They compute the displacements of different sets of points along the sequence of frames describing each facial expression (i.e. from neutral to apex). First, we evaluate our algorithms on the Cohn-Kanade database for the six prototypic expressions under two different spatial frame resolutions (original and 40%-reduced). Later, our methods were also tested on the MMI database which presents higher variabilities than the Cohn-Kanade one. The results on the first database show that dense flow tracking method at original resolution slightly outperformed, in average, the recognition rates of feature point tracking method (95.45% against 92.42%) but it requires 68.24% more time to track the points. For the patterns of MMI database, using dense flow tracking at the original resolution, we achieved very similar average success rates.     

OpenCV分水岭算法的改进及其在细胞分割中的应用

分水岭算法是一种基于形态学的图像分割算法,能快速准确地确定连通区域的边界.将基于标记的分水岭算法用于细胞图像的分割,较好地解决了粘连细胞的分割问题.在该细胞分割算法的实现过程中,发现了OpenCV分水岭算法实现的缺陷,通过对相关代码的分析,发现该缺陷存在的原因是算法流程中对相邻像素相对关系的描述存在问题.将OpenCV分水岭算法中对相邻像素取差的绝对值,改为对相邻像素取差值,对该算法进行了改进.实验证明,改进后的OpenCV分水岭算法对细胞图像的分割效果明显好于直接使用OpenCV分水岭算法得到的结果.该方法在不影响分割速度的情况下,提高了OpenCV分水岭算法分割的准确度.     

A study on the possibility of implementing a real-time stereoscopic 3D rendering TV system

3D video has recently seen a massive increase in exposure in our lives. However, differences between the viewing and shooting conditions for a film lead to disparities between the reformed media and the original three-dimensional effect, which cause severe visual fatigue to viewers and result in headaches and dizziness. In this paper, a series of image processing algorithms are introduced to overcome these problems. The image processing pipeline is composed of four steps, eye-pupil detection, stereo correspondence computation, saliency map generation, and 3D warping. Each step is implemented in an S3DS-3D rendering system and its time complexity is measured. From the results, it was found that real-time stereoscopic 3D rendering is impossible using only a software implementation because SIFT and optical flow calculation requires a significant amount of time. Therefore, these two algorithm blocks should be implemented with hardware acceleration. Fortunately, active research is being conducted on these issues and real-time processing is expected to become available soon for applications beyond full-HD TV screens. In addition, it was found that saliency map generation and 3D warping blocks also need to be implemented in hardware for full-HD display although they do not have significant time complexity compared to SIFT and optical flow algorithm blocks.     

融合CamShift的TLD算法实现人脸跟踪

针对跟踪-学习-检测(Tracking-Learning-Detection,TLD)算法跟踪模块所用金字塔光流法计算量大,跟踪人脸实时性差的问题,提出融合连续自适应均值漂移(Continuously Adaptive Mean Shift,CamShift)的TLD算法提高人脸跟踪效率.改进的TLD算法框架中跟踪模块选用CamShift算法实现目标人脸跟踪,检测模块采用滑动窗法扫描搜索,再使用分类器判断目标是否存在,学习模块根据跟踪模块和检测模块的结果对比评估错误和误差,更新目标模型.将改进的TLD算法分别与CamShift算法和TLD算法进行对比试验,结果表明,融合CamShift的TLD算法实现人脸跟踪效率和准确率均高于原始两种算法,且满足实时性要求.     

基于密集轨迹和光流二值化图的行为识别算法

针对改进的密集轨迹算法（improved dense trajectories,iDT）提取的轨迹数量较为庞大的问题,提出了一种轨迹滤除方法。密集采样兴趣点,利用光流图计算每个兴趣点下一帧的位置进而组成轨迹,对每帧光流图进行最大值归一化以及二值化,得到光流二值化图,以此反映该点的运动是否相对显著。利用光流二值化图统计轨迹上各点的有效性从而判断轨迹是否满足有效条件,并将不满足条件的轨迹滤除,得到提纯的轨迹。为了验证算法的有效性,使用了行为识别领域的常用数据集KTH和UCF sports对算法进行验证,实验结果表明,该算法能在保证准确率的同时减少轨迹数量,并且计算量较小。