文章目录

project(ch8)set(CMAKE_BUILD_TYPE

Release)

-marchnative)find_package(OpenCV

REQUIRED)

REQUIRED)include_directories(${OpenCV_INCLUDE_DIRS}${G2O_INCLUDE_DIRS}${Sophus_INCLUDE_DIRS}/usr/include/eigen3/${Pangolin_INCLUDE_DIRS}

optical_flow.cpp)

target_link_libraries(optical_flow

${OpenCV_LIBS})#[[

target_link_libraries(direct_method

${OpenCV_LIBS}

/home/xixi/Downloads/slambook2-master/ch8/optical_flow.cpp:145:37:

error:

public:OpticalFlowTracker(const

Mat

inverse(inverse_),has_initial(has_initial_)

{}void

success_single;OpticalFlowSingleLevel(img1,

img2,

success_multi;chrono::steady_clock::time_point

chrono::steady_clock::now();OpticalFlowMultiLevel(img1,

img2,

true);chrono::steady_clock::time_point

chrono::steady_clock::now();auto

time_used

chrono::duration_castchrono::durationdouble(t2

t1);cout

pt1.push_back(kp.pt);vectoruchar

status;vectorfloat

chrono::steady_clock::now();cv::calcOpticalFlowPyrLK(img1,

img2,

chrono::steady_clock::now();time_used

chrono::duration_castchrono::durationdouble(t2

t1);cout

img2_single);cv::imshow(tracked

multi

{kp2.resize(kp1.size());success.resize(kp1.size());OpticalFlowTracker

tracker(img1,

has_initial);parallel_for_(Range(0,

kp1.size()),std::bind(OpticalFlowTracker::calculateOpticalFlow,

tracker,

OpticalFlowTracker::calculateOpticalFlow(const

Range

pyramidschrono::steady_clock::time_point

chrono::steady_clock::now();vectorMat

pyr1,

{pyr1.push_back(img1);pyr2.push_back(img2);}

else

pyramid_scale));cv::resize(pyr2[i

1],

pyramid_scale));pyr1.push_back(img1_pyr);pyr2.push_back(img2_pyr);}}chrono::steady_clock::time_point

chrono::steady_clock::now();auto

time_used

chrono::duration_castchrono::durationdouble(t2

t1);cout

1];kp1_pyr.push_back(kp_top);kp2_pyr.push_back(kp_top);}for

(int

chrono::steady_clock::now();OpticalFlowSingleLevel(pyr1[level],

pyr2[level],

chrono::steady_clock::now();auto

time_used

chrono::duration_castchrono::durationdouble(t2

t1);cout

稀疏直接法关键点假设周围像素不变(因此不必计算描述子)。

可快速求解相机位姿。

稠密直接法多数需要GPU加速

/home/xixi/Downloads/slambook2-master/ch8/direct_method.cpp:206:35:

error:

CMakeFiles/direct_method.dir/direct_method.cpp.o:

function

JacobianAccumulator::accumulate_jacobian(cv::Range

const):

/home/xixi/Downloads/slambook2-master/ch8/direct_method.cpp:235:

undefined

Sophus::SE3::operator*(Eigen::Matrixdouble,

const)

CMakeFiles/direct_method.dir/direct_method.cpp.o:

function

DirectPoseEstimationSingleLayer(cv::Mat

const,

std::vectorEigen::Matrixdouble,

Eigen::aligned_allocatorEigen::Matrixdouble,

const,

/home/xixi/Downloads/slambook2-master/ch8/direct_method.cpp:178:

undefined

Sophus::SE3::exp(Eigen::Matrixdouble,

const)

/home/xixi/Downloads/slambook2-master/ch8/direct_method.cpp:178:

undefined

Sophus::SE3::operator*(Sophus::SE3

const)

target_link_libraries(direct_method

${OpenCV_LIBS}

${Sophus_LIBRARIES})byzanz-record

146

/home/xixi/myGIF/test.gif这里程序运行感觉不太对暂时不清楚哪里。

direct_method.cpp

Eigen::aligned_allocatorEigen::Vector2d

VecVector2d;//

public:JacobianAccumulator(const

cv::Mat

DirectPoseEstimationMultiLayer(const

cv::Mat

DirectPoseEstimationSingleLayer(const

cv::Mat

depthdepth_ref.push_back(depth);pixels_ref.push_back(Eigen::Vector2d(x,

y));}//

DirectPoseEstimationSingleLayer(left_img,

img,

T_cur_ref);DirectPoseEstimationMultiLayer(left_img,

img,

DirectPoseEstimationSingleLayer(const

cv::Mat

chrono::steady_clock::now();JacobianAccumulator

jaco_accu(img1,

{jaco_accu.reset();cv::parallel_for_(cv::Range(0,

px_ref.size()),std::bind(JacobianAccumulator::accumulate_jacobian,

jaco_accu,

std::placeholders::_1));Matrix6d

jaco_accu.bias();//

chrono::steady_clock::now();auto

time_used

chrono::duration_castchrono::durationdouble(t2

t1);cout

cv::COLOR_GRAY2BGR);VecVector2d

projection

jaco_accu.projected_points();for

(size_t

JacobianAccumulator::accumulate_jacobian(const

cv::Range

half_patch_size)continue;projection[i]

Eigen::Vector2d(u,

J_pixel_xi).transpose();hessian

J.transpose();bias

DirectPoseEstimationMultiLayer(const

cv::Mat

{pyr1.push_back(img1);pyr2.push_back(img2);}

else

pyramid_scale));cv::resize(pyr2[i

1],

pyramid_scale));pyr1.push_back(img1_pyr);pyr2.push_back(img2_pyr);}}double

fxG

{px_ref_pyr.push_back(scales[level]

px);}//

scales[level];DirectPoseEstimationSingleLayer(pyr1[level],

pyr2[level],

强烈非凸。

优化算法易进入极小只有运动很小时直接法才能成功。

金字塔的引入可以在一定程度上减小非凸的影响。

2、单个像素无区分度

Search稠密逆搜索)光流算法【低时间复杂度有竞争力的精度】

DIS光流算法。

这个类实现了密集逆搜索(DIS)光流算法。

包括三个预设带有预选参数在速度和质量之间提供合理的权衡。

但是即使是最慢的预设也还是比较快的如果你需要更好的质量不关心速度可以使用DeepFlow。

Till

【基于翘曲理论的高精度光流估计角误差更小对参数变化不敏感噪声鲁棒】Thomas

Brox,

将一个连续的、旋转不变的能量泛函用于光流计算该泛函基于两个项:一个具有亮度常数和梯度常数假设的鲁棒数据项结合一个保持不连续的时空

cv::VariationalRefinement::calcUV()

稀疏光流

该类可以使用金字塔迭代Lucas-Kanade方法计算稀疏特征集的光流。

u-1

方向上的梯度值。

这种做法有什么缺点提示对于距离较近的特征变化应该较快而距离较远的特征在图像中变化较慢求梯度时能否利用此信息

3、直接法是否能和光流一样提出“反向法”的概念即使用原始图像的梯度代替目标图像的梯度

单目直接法在优化时