Study on Local Optical Flow Method Based on YOLOv3 in Human Behavior Recognition

In the process of human behavior recognition, the traditional dense optical flow method has too many pixels and too much overhead, which limits the running speed. This paper proposed a method combing YOLOv3 (You Only Look Once v3) and local optical flow method. Based on the dense optical flow method, the optical flow modulus of the area where the human target is detected is calculated to reduce the amount of computation and save the cost in terms of time. And then, a threshold value is set to complete the human behavior identification. Through design algorithm, experimental verification and other steps, the walking, running and falling state of human body in real life indoor sports video was identified. Experimental results show that this algorithm is more advantageous for jogging behavior recognition.

An Improved Optical Flow Method for Image Registration with Large-scale Movements

在这份报纸，为图象登记的一个改进光流动方法被建议。它以它与一个起始的运动评估者改进光流动方法的方法是新奇的：扩大了阶段关联技术(EPCT ) ，用补偿前者的缺乏的后者的优点。以一种更多的详细方式，光流动方法能到达亚象素精确性并且计算象啁啾并且倾斜一样的复杂失真模式，这能被说但是与大规模运动弱。因为 EPCT 用象素水平精确性盖住大翻译和旋转的大小并且在精明的负担是有效的，它能为光流动方法被当作一个好起始的运动评估者。测试证明了特别，这个改进方法将显著地与大规模运动为图象提高登记表演并且对随机的噪音柔韧。

Three-dimensional color particle image velocimetry based on a cross-correlation and optical flow method

Rainbow particle image velocimetry(PIV)can restore the three-dimensional velocity field of particles with a single camera;however,it requires a relatively long time to complete the reconstruction.This paper proposes a hybrid algorithm that combines the fast Fourier transform(FFT)based co-correlation algorithm and the Horn–Schunck(HS)optical flow pyramid iterative algorithm to increase the reconstruction speed.The Rankine vortex simulation experiment was performed,in which the particle velocity field was reconstructed using the proposed algorithm and the rainbow PIV method.The average endpoint error and average angular error of the proposed algorithm were roughly the same as those of the rainbow PIV algorithm;nevertheless,the reconstruction time was 20%shorter.Furthermore,the effect of velocity magnitude and particle density on the reconstruction results was analyzed.In the end,the performance of the proposed algorithm was verified using real experimental single-vortex and double-vortex datasets,from which a similar particle velocity field was obtained compared with the rainbow PIV algorithm.The results show that the reconstruction speed of the proposed hybrid algorithm is approximately 25%faster than that of the rainbow PIV algorithm.

An improved optical flow method to estimate Arctic sea ice velocity(winter 2014-2016)

Sea ice velocity impacts the distribution of sea ice,and the flux of exported sea ice through the Fram Strait increases with increasing ice velocity.Therefore,improving the accuracy of estimates of the sea ice velocity is important.We introduce a pyramid algorithm into the Horn-Schunck optical flow(HS-OF)method(to develop the PHS-OF method).Before calculating the sea ice velocity,we generate multilayer pyramid images from an original brightness temperature image.Then,the sea ice velocity of the pyramid layer is calculated,and the ice velocity in the original image is calculated by layer iteration.Winter Arctic sea ice velocities from 2014 to 2016 are obtained and used to discuss the accuracy of the HS-OF method and PHS-OF(specifically the 2-layer PHS-OF(2 LPHS-OF)and 4-layer PHS-OF(4 LPHS-OF))methods.The results prove that the PHS-OF method indeed improves the accuracy of sea ice velocity estimates,and the 2 LPHS-OF scheme is more appropriate for estimating ice velocity.The error is smaller for the 2 LPHS-OF velocity estimates than values from the Ocean and Sea Ice Satellite Application Facility and the Copernicus Marine Environment Monitoring Service,and estimates of changes in velocity by the 2 LPHS-OF method are consistent with those from the National Snow and Ice Data Center.Sea ice undergoes two main motion patterns,i.e.,transpolar drift and the Beaufort Gyre.In addition,cyclonic and anticyclonic ice drift occurred during winter 2016.Variations in sea ice velocity are related to the open water area,sea ice retreat time and length of the open water season.

Internal Flow Field Measurement of Gas Turbine Based on Optical Flow Method

The in-cylinder flow field of the internal combustion engine is an important factor affecting the quality and combustion quality of the fuel mixture in the cylinder. In order to calculate the high-precision flow field, the paper presents a flow field calculation method based on the optical flow algorithm. The motion of the point was calculated using the change in pixel intensity within two temporally adjacent frame images. The results show the high accuracy and resolution of the flow field at small displacement conditions.

Discontinuity-preserving optical flow algorithm

A modification of Horn and Schunk＇s approach is investigated, which leads to a better preservation of flow discontinuities. It improves Horn-Schunk model in three aspects： （1） It replaces the smooth weight coefficient in the energy equation by the variable weight coefficient. （2） It adopts a novel method to compute the mean velocity. The novel method also reflects the effect of the intensity difference on the image velocity diffusion. （3） It introduces a more efficient iterative method than the Gauss-Seidel method to solve the associated Euler-Lagrange equation. The experiment results validate the better effect of the improved method on preserving discontinuities.

Comparision of DdQm models in image optical flow analysis based on LBM

The optical flow analysis of the image sequence based on the formal lattice Boltzmann equation, with different DdQm models, is discussed in this paper. The Mgorithm is based on the lattice Boltzmann method （LBM）, which is used in computational fluid dynamics theory for the simulation of fluid dynamics. At first, a generalized approximation to the formal lattice Boltzmann equation is discussed. Then the effects of different DdQm models on the results of the optical flow estimation are compared with each other, while calculating the movement vectors of pixels in the image sequence. The experimental results show that the higher dimension DdQm models, e.g., D3Q15 are more effective than those lower dimension ones.

A rain-type adaptive optical flow method and its application in tropical cyclone rainfall nowcasting

A rain-type adaptive pyramid Kanade-Lucas-Tomasi(A-PKLT)optical flow method for radar echo extrapolation is proposed.This method introduces a rain-type classification algorithm that can classify radar echoes into six types:convective,stratiform,surrounding convective,isolated convective core,isolated convective fringe,and weak echoes.Then,new schemes are designed to optimize specific parameters of the PKLT optical flow based on the rain type of the echo.At the same time,the gradients of radar reflectivity in the fringe positions corresponding to all types of rain echoes are increased.As a result,corner points that are characteristic points used for PKLT optical flow tracking in the surrounding area will be increased.Therefore,more motion vectors are purposefully obtained in the whole radar echo area.This helps to describe the motion characteristics of the precipitation more precisely.Then,the motion vectors corresponding to each type of rain echo are merged,and a denser motion vector field is generated by an interpolation algorithm on the basis of merged motion vectors.Finally,the dense motion vectors are used to extrapolate rain echoes into 0-60-min nowcasts by a semi-Lagrangian scheme.Compared with other nowcasting methods for four landfalling typhoons in or near Shanghai,the new optical flow method is found to be more accurate than the traditional cross-correlation and optical flow methods,particularly showing a clear improvement in the nowcasting of convective echoes on the spiral rainbands of typhoons.

Experiment on Vaporization of Jet into Cross-Flow

The injection characteristics of the main fuel nozzle,which is widely applied in advanced lean-premixed-prevaporized(LPP)low-emission combustors,can be simplified as the atomization and vaporization processes of a jet into cross-flow.In this study,a nozzle with a diameter of 0.4 mm is designed and processed through the heating of the inlet air,and the vaporization characteristics are investigated.The optical measurement and cyclone separation methods are separately used to investigate the evaporation rate of a jet into cross-flow.Experimental results show that the fuel evaporation rate in cross-flow is mainly affected by the Weber number(We),equivalent ratio(φ),momentum rate of fuel to air(q),and air temperature.In addition,the inlet temperature is a crucial factor for the evaporation ratio of a jet into cross-flow.The evaporation results measured by two different methods in the same cross-flow are very close to each other with a deviation within 10%.

基于双目深度筛选的ORB-SLAM3算法

针对ORB-SLAM3算法中特征点存在易丢失、精度低,进而导致双目在复杂场景下运动轨迹误差大的问题,本文设计了一种改进的ORB-SLAM3算法。首先,在ORB特征匹配算法中引入自适应角点检测技术,增加特征点的采集数量,并采用光流法跟踪图像特征,提高关键帧的创建成功率;其次,以特征点为中心,作区域搜索,提高实时性;然后,采用双向左右一致性检验筛选最优视差,应用Prosac算法去除误匹配点对;最后,结合深度信息对关键帧进行筛选,提高关键帧的质量,优化相机位姿。采用KITTI和EuRoc数据集进行了试验,验证了改进算法在绝对轨迹误差上具有良好的优化效果。

基于视觉与加速度测量的结构动态位移融合估计

结构动态位移测量与精准估计对于结构安全运营和性态评估具有重要意义.基于计算机视觉的位移监测方法具有精度高、非接触式、成本低、设备安装简便等优点,在实际复杂工程环境中,设备难以架设,视觉测量方法较传统接触式位移监测方法具有明显的优势.图像分辨率和拍摄帧率等因素在一定程度上限制了视觉方法的使用.针对视觉位移测量技术高频振动识别精度低的问题,提出了一种基于视觉与加速度测量的结构动态位移重构方法,通过融合视觉低频与加速度高频振动响应信号,实现结构动态位移精准识别.首先,利用光流法从结构振动视频数据中提取结构位移响应,引入前后向误差与离群值过滤机制,提升特征点追踪精度,避免漂移问题.然后,利用逐次变分模态分解方法分别从视觉位移与加速度二次积分得到的位移信号中提取相应的本征模态函数(intrinsic mode function,IMF)分量.最后,基于互相关函数筛选机制,确定融合模态分量,融合基于视觉测量的低阶IMF与基于加速度测量的高阶IMF,重构结构位移响应.通过一个钢筋混凝土框架结构振动台试验,对提出的位移融合估计方法进行了试验验证.结果表明:与单一视觉测量方法相比,所提出的方法能够更为准确地估计结构动态位移,并且通过引入加速度测量中的动态位移分量,融合后的位移比基于视觉测量的结果具有更宽的频率范围.

基于光流法的流体运动预测技术适用性研究

由于粒子图像测速技术在测量精度和时空分辨率上难以同时保证,而基于光流法的流体运动估计方法可以获得具有较高分辨率的稠密速度矢量场。为了评价该方法在流体运动估计上的适用性,采用流动情况由简单到复杂的开源粒子图像、标量图像、纹影图像序列对3种经典光流算法进行实验,在定性和定量层面评估了各算法的表现。结果表明:传统HS光流算法适用于大部分流动情况,然而在灰度梯度变化不明显时表现不佳;基于随机湍流传输模型的变分光流算法表现更为稳定,在HS目标函数中添加加权非局部平滑项后更适用于捕捉涡轮叶片纹影图像的尾迹区域和估计尾迹涡的输运速度。

基于改进卷积神经网络的心理状态预警技术

针对传统问卷法难以真实反映被调查者心理状态的问题,基于光流法和卷积神经网络提出了一种微表情判断方法,并将其作为心理状态预警技术的核心模块。对于数据集中人脸数据离散的问题,该方法采用人眼权重法对图像进行预处理,且通过金字塔光流算法提取预处理图像序列的光流特征,再利用三维卷积神经网络对该特征加以训练。与传统算法相比,所提方法在减少模型训练参数与运算时间的同时还具有更优的学习能力。实验测试结果表明,该算法在CASME数据集上的微表情识别准确率为89.2%,F1值为0.6751,均优于其他对比方法。由此证明,该算法可实现对人脸微表情的准确识别,进而为学生心理状态预警提供客观的数据支撑。

基于数据集蒸馏的光伏发电功率超短期预测

云是影响太阳直接辐射变化的主要因素,由于各类云的透光率不同,导致到达光伏电站的太阳辐射会随之产生波动。为解决各类云遮挡下的光伏发电功率波动大、预测模型个数多的问题,提出一种基于卫星云图和数据集蒸馏的光伏发电功率超短期预测模型。首先,基于待测场站上方的历史云图,采用Farneback光流法预测出云图;然后,根据卫星云分类标签数据建立各类云的样本库,利用数据集蒸馏算法训练样本库得到云类判别图,将预测云图与云类判别图匹配计算,获得云类聚合匹配特征;最后,利用上述特征、云量特征以及数值天气预报数据建立长短期记忆网络模型,对光伏发电功率进行超短期预测。利用某光伏电站数据进行验证,结果显示,该文所提模型能准确描述云层的各项特征,有效提升光伏功率预测精度。

动态场景下基于YOLOv5和几何约束的视觉SLAM算法

目的移动智能体在执行同步定位与地图构建(Simultaneous Localization and Mapping,SLAM)的复杂任务时,动态物体的干扰会导致特征点间的关联减弱,系统定位精度下降,为此提出一种面向室内动态场景下基于YOLOv5和几何约束的视觉SLAM算法。方法首先,以YOLOv5s为基础,将原有的CSPDarknet主干网络替换成轻量级的MobileNetV3网络,可以减少参数、加快运行速度,同时与ORB-SLAM2系统相结合,在提取ORB特征点的同时获取语义信息,并剔除先验的动态特征点。然后,结合光流法和对极几何约束对可能残存的动态特征点进一步剔除。最后,仅用静态特征点对相机位姿进行估计。结果在TUM数据集上的实验结果表明,与ORB-SLAM2相比,在高动态序列下的ATE和RPE都减少了90%以上,与DS-SLAM、Dyna-SLAM同类型系统相比,在保证定位精度和鲁棒性的同时,跟踪线程中处理一帧图像平均只需28.26 ms。结论该算法能够有效降低动态物体对实时SLAM过程造成的干扰,为实现更加智能化、自动化的包装流程提供了可能。

基于视觉增强的桥梁检测算法及技术

为了识别桥梁结构微小振动,提出一种基于方向码匹配(OCM)法的靶标圆心检测算法.该算法首先采用OCM法实现像素级匹配,然后在最佳匹配区域采用靶标圆心检测算法进行精确跟踪.通过室内试验,将基于OCM的靶标圆心检测算法与基于相位的光流法和OCM亚像素实现算法进行对比,研究3种算法在不同振动幅度和实际工程应用的适用性,为实桥检测提供技术方案比选依据.结果表明:所提算法在各种振动幅度下均能精确地识别结构振动位移和频率,在实桥应用时应考虑靶标粘贴的可行性;基于相位的光流法在幅度指标小于2 mm或1.7像素的微小振动工况下能精确地识别结构的振动位移和频率;OCM亚像素实现算法的分辨精度取决于插值点的数量,对大幅度振动工况的识别具有与所提算法相似的效果.

光流法雷达外推产品在突发强降水预报中的应用

地形条件复杂的山丘区中小河流洪水突发性强、汇流时间短,高准确率和长时效性降水短时临近预报产品对提高突发洪水预报精度尤为关键。以2021年9月河南省鸭河口水库出现的千年一遇特大洪水为例,利用国家气象信息中心提供的三源融合格点实况降水资料,检验基于改进光流法的雷达外推降水预报产品在本次洪水过程中0~1 h和0~3 h降水预报的TS评分和预报偏差。结果表明:(1)改进光流法在0~1 h的逐小时降水预报上较为精准,累计雨量在50 mm以下时,TS评分在0.45~0.85之间;雨量在50~70 mm之间时,TS评分在0.35~0.70之间;雨量在70 mm以上时,TS评分在0.25~0.35之间。50 mm以上雨量有较高TS评分表现出改进光流法在极端强降水预报中的优势性。(2)改进光流法在0~3 h的降水预报上,累计雨量在50 mm以下时,TS评分在0.55~0.85之间;在50 mm以上时,TS评分在0.35~0.75之间。该降水预报产品不仅对极端性降水预报效果较好,且预报时效长达3 h,可为防洪调度提供更长的决策时间。(3)改进光流法在0~3 h的降水预报产品与融合实况格点降水相比,雨量在20 mm以下的预报结果比较接近,平均绝对误差在10 mm以内;雨量在20 mm以上时,随雨量增大,平均误差、平均绝对误差、均方根误差均逐渐增大。(4)改进光流法在0~3 h的降水预报产品对影响范围小、降水强度大、维持时间长、累计雨量大的极端强降水有较好的预报表现。研究成果可为洪水预报模型提供一种较为可靠的降水输入预报。

基于金字塔分层L-K光流法的燃气火焰动态特征提取与分析

本文研究了金字塔分层L-K光流算法在燃气火焰诊断中的应用,通过实验提取不同工况下燃气火焰图像动态特征,并验证火焰动态特征灵敏性和有效性.结果表明:在稳定工况下,速度大小的均值和方差相比于速度方向的均值和方差更能反映火焰的动态变化;在不稳定工况下,动态特征的速度大小、均值出现周期性变化,描述了不稳定燃烧和火焰熄灭的过程;速度大小均值和方差与工况参数之间呈现显著相关,而速度方向均值和方差相关性较弱.本研究为开发新的燃气检测技术提供参考.

基于无人机航拍视频的运动目标检测与跟踪

针对无人机视角下拍摄物体尺寸小、外观模糊,以及无人机平台自身存在运动等问题,基于无人机航拍视频的运动目标检测与跟踪仍具有很大的挑战性。提出了基于改进LK光流法的运动目标检测与跟踪算法,利用卡尔曼滤波算法进行目标跟踪;针对传统LK光流法在动态环境下效果不佳的问题,引入语义分割线程筛选动态点和静态点,通过筛选后的静态点求解单应矩阵完成无人机的运动补偿、计算动态点的光流值获得光流点以降低误匹配率,最后通过聚类算法和形态学运算得到完整的运动目标;将所提的重识别模块引入卡尔曼滤波算法实现目标跟踪。实验结果验证了所设计的运动目标检测与跟踪算法在无人机视角下能够实时、准确地提取出运动目标,并持续稳定地跟踪目标。

隧洞平行摄影测量立体影像尺度差估计及匹配优化

为提升平行摄影隧洞序列影像的同名点匹配率及精度,对成像规律及匹配方法展开研究。提出了一种窗口尺度自动改正的LK光流法,核心是利用物像空间关系构建尺度差模型,并采用可变窗口进行特征点光流跟踪。结果表明:隧洞序列影像像点尺度差呈对称性径向分布,差异值符合幂函数模型增长趋势;将该方法应用在多组尺度差立体像对的匹配实验中,均取得了优于0.3个像素精度的实验结果,较基本光流法至少提升34.3%,最大提升45.5%。研究成果可为顾及尺度差的匹配方法提供参考,也可为隧洞平行摄影立体影像匹配提供基础。