Copyright By PowCoder代写 加微信 powcoder
Calibration
Least squares line fitting
Hought transform
点到拟合直线y方向的距离，平方求和
not rotation- invariant
fails for vertical line 平行于y轴的直线找不到y方向距离
total least suares
点到拟合直线的最短距离，平方求和
prons robust to rotation
cons not robust to outliers 噪点到直线的距离很大，平方后更大
deal with multiplu RANSAC
choose a small subset of points uniformally at random
fit a model to that subset
put all the points below a distance thredshold into the subset
number of trial S
number of sampled point K minimum number needed to fit the model
distance threshold
起始的点的个数？ determined by a fomula
not good for getting multiple fits
lost of parameters to tune
image stitching using affine transform
estimate affine transformation T with k random selected match points
if , preseve T
choose the best T and recalculate the transform image
a line mapping
to Hough space
直线映射成点，点映射成直线
在xy平面拟合直线即houghspace求两直线线交点y=ax+b to (a,b)
problem not able to model a vertical line
use quantized
make a table to count how many times each pair appears
dealing with noisefind local maximum
a line mapping to Hough space
1.根据(x,y)和角度 画图表，写出对应的r 角度范围-45~90
2. 根据上一个图标，画角度和r的图标，写出每对出现的次数
3. 找出出现次数最多的一对，写出直线表达式
thresholding-based
segmentation
single thresholding
double thresholding
watch out the histogram to predict appropriate threshold
Otsu’s method
Optimum Global Thresholding
minimize with-in class variance
adaptive thresholding using moving adverages
maixmize between class variance
T = b * local mean
k-mean clustering
randomly initialize the cluster center
determine points in each cluster
recalcualate the cluster center
group based on
RGB/colour
lips RGB+position
斑点裙子 texture
prons fast
sensitive to
initial centers
detect spherical clusters
transformation from
world coordinate to
image coordinate
1. world coordinate to
camera/normalized coordinata
canera center is origin point
principal axis is z-axis
x,y-axis are parallel to x and y axes of world coordinate
principal point point where principal axis intersects the image plane
2. camera coordinate
to image coordinate origin is in the center
world to camera
camera to image
R should be three normallized column vectors 并且相互垂直
C is coords. of camera
center in world frame
P=K[R|t] (t=-RC)
intrinsic parameter (K)
pixel coordinates
pixels per meter in horizontal/vertica ldirection
principal point coordinate
focal length
pixel magnification factors
extrinsic parameter (R,t)
roatation and translation matrix
camera to image
world to camera
measure height
calcualation
know world to image coordinate pairs to estimate matrix minimal for 6 pair
know matrix and world/image coordinate to estimate image/world coordinate
calibration for
vanishing points
one vanishing point
cannot solve focal length
principal point is the vanishing point
two or three vanishing point can solve for focal length and principal point
epipolar geometry
baseline line connecting the two camera centre
epipolar plane plane containing baseline
intersection of basline with image plane
projections of the other camera center
epipolar lines
intersections of epipolar plane with image plane
one line maps to another line
pipline epipolar rectification stereo matching trangulation
parallel images
images planes of cameras are parallel to each other and to the baseline
camera centers are at the same height
focal lengths are the same
epipolar retifaction
to achieve parallel images
block matching only apply on the horizontal scan lines which is epipolar line
1. roatate the right camera by R
2. rotate the left camera so that epipole is at infinity
3. so rotate the right camera so that the epipole is at infinity
4. adjust the scale
stereo matching
find correspoding points in epipolar line
1. slide window along the right scanline and compare two window
2. matching cost and find the disparity at the maximum
point(if normalized correlation) or minimum point(if SSD)
triangulation
disparity is invesely propotinal to depth
window-based
1. matching cost computation
2. cost aggregation
3. disparity computation
local base method
window size
small more detail but more noise
large smoother disparity maps but less details
untertured region problem
aperture problem
texture with only horizontal orientation
needs colour varitation inside
texture with vertical orientation
multiple fit
points problem
repetive pattern
foreground fattening problem
more obvious for large window size
because of cost aggregation
adaptive support
only pixels that lie on the same disparity contribute to the aggregated matching costs
cost aggregation
c for color distance, s for spatial distance
large distance for low weight
similar to bilateral filter
non local constraints
uniqueness for one point, only one matching point
ordering corresponding points hould be in the same order
smooth expect disparity values to change slowly
calcualte depth from disparity
程序代写 CS代考 加微信: powcoder QQ: 1823890830 Email: powcoder@163.com