def estimate(self, first_set, second_set):
set_length = len(first_set)
best_model = None
best_model_inliers = 0
inliers_indices = []
range_set_length = range(set_length)
for _ in xrange(self.max_iterations):
if set_length < 3:
continue
random_indices = None
if set_length > self.hypothesis_set_length:
# This will return a list of hypothesis_set_length numbers selected from the range 0 to set_length, without duplicates.
random_indices = random.sample(range_set_length, self.hypothesis_set_length)
else:
random_indices = range_set_length
# create random subset
first_subset = np.array([first_set[i] for i in random_indices], dtype=np.float32)
second_subset = np.array([second_set[i] for i in random_indices], dtype=np.float32)
# estimate model on random subset
current_model = cv2.estimateRigidTransform(first_subset, second_subset, fullAffine=False)
if current_model is None:
continue
current_model_inliers = 0
current_inliers_indices = []
# count inliers
for index in xrange(set_length):
transformed_point = transform(current_model, first_set[index][0])
error = math.sqrt(
math.pow(transformed_point[0] - second_set[index][0][0], 2)
+ math.pow(transformed_point[1] - second_set[index][0][1], 2)
)
if error < self.max_distance:
current_model_inliers += 1
if self.remember_inlier_indices:
current_inliers_indices.append(index)
if current_model_inliers > best_model_inliers:
best_model = current_model
best_model_inliers = current_model_inliers
inliers_indices = current_inliers_indices
if best_model is None or (self.min_inliers is not None and best_model_inliers < self.min_inliers):
best_model = cv2.estimateRigidTransform(first_set, second_set, fullAffine=False)
if self.remember_inlier_indices:
inliers_indices = [i for i in xrange(set_length)]
return best_model, inliers_indices
def RANSAC_2D(point_set1, point_set2, iteration, tolerance):
x_mean = np.mean(point_set2[:,0])
y_mean = np.mean(point_set2[:,1])
print x_mean
print y_mean
point_count = point_set1.shape[0]
best_M = None
best_inlier = 0
best_model = None
for i in range(0, iteration):
sample_index = np.random.choice(range(0, point_count), 3, replace = False)
first_set = np.array([point_set1[sample_index]], dtype = np.float32)
second_set = np.array([point_set2[sample_index]], dtype = np.float32)
#print first_set.shape
transformation = cv2.estimateRigidTransform(first_set, second_set, fullAffine = False)
if transformation is None:
continue
#print '******'
#print first_set
#print second_set
#print '******'
#transformation[1,2] *= -1
new_point_set1 = np.asarray([T2D.Mirror(transformation, x, x_mean, y_mean) for x in point_set1], dtype=np.float32)
distance_matrix = np.linalg.norm(new_point_set1 - point_set2, axis = 1)
#print distance_matrix
inlier = np.sum(distance_matrix <= tolerance)
#print inlier
if inlier > best_inlier:
best_inlier = inlier
best_M = transformation
best_model = new_point_set1
print best_inlier
print np.linalg.pinv(best_M[:, 0:2])
return best_M, best_model
def homography_transform(rotated_images1, rotated_images2):
dst1 = rotated_images1
dst2 = rotated_images2
h, w = dst1.shape[:2]
#compute matching features from two image pairs
image1_pts, image2_pts = extract_and_match(dst1, dst2)
# #code to output the features matched in each image
# image = cv2.addWeighted(rotated_images1,0.5,rotated_images2,0.5,0)
# image = draw_matches(image, image1_pts, image2_pts)
# cv2.imshow('image', image)
# cv2.waitKey(0)
# cv2.destroyAllWindows()
# sys.exit()
print(image1_pts)
print(image2_pts)
n = len(image1_pts)
image2_pts_np = np.reshape(np.float32(image2_pts),(1,n,2))
image1_pts_np = np.reshape(np.float32(image1_pts),(1,n,2))
#compute homography transform from img1 to img2
if(len(image1_pts)>=4):
homography, mask = cv2.findHomography(np.float32(image1_pts), np.float32(image2_pts), cv2.RANSAC, 1.0)
affine = cv2.estimateRigidTransform(np.float32(image1_pts), np.float32(image2_pts), fullAffine=True)
print(homography)
return homography, affine
else:
return None, None
def align_5_points_eye_center(self, imgDim, rgbImg, five_points):
'''
@brief 根据：两个眼睛中心，两个嘴角，一个鼻子这五点进行对齐,其它与align_dlib_cpp保持一致。
@param five_points: 一个shape 为10 的人脸关键点坐标位置。
@attention 适合于CelebA提供的数据。
'''
assert imgDim is not None
assert rgbImg is not None
from_points = []
for i in range(0,10,2):
from_points.append((five_points[i],five_points[i+1]))
to_points = []
five_mean_x = [(self.mean_shape_x[36-17]+self.mean_shape_x[39-17])*0.5,(self.mean_shape_x[42-17]+self.mean_shape_x[45-17])*0.5,self.mean_shape_x[33-17],self.mean_shape_x[48-17],self.mean_shape_x[54-17]]
five_mean_y = [(self.mean_shape_y[36-17]+self.mean_shape_y[39-17])*0.5,(self.mean_shape_y[42-17]+self.mean_shape_y[45-17])*0.5,self.mean_shape_y[33-17],self.mean_shape_y[48-17],self.mean_shape_y[54-17]]
for i in range(5):
new_ref_x = (self.padding+five_mean_x[i])/(2*self.padding+1)
new_ref_y = (self.padding+five_mean_y[i])/(2*self.padding+1)
to_points.append((imgDim *new_ref_x,imgDim *new_ref_y))
source = np.array(from_points).astype(np.int)
target = np.array(to_points,).astype(np.int)
source = np.reshape(source,(1,5,2))
target = np.reshape(target,(1,5,2))
H = cv2.estimateRigidTransform(source,target,False)
if H is None:
return
else:
aligned_face = cv2.warpAffine(rgbImg, H, (imgDim, imgDim))
return aligned_face
def align_5_points(self, imgDim, rgbImg, five_points):
'''
@brief 根据：两个眼角，两个嘴角，一个鼻子这五点进行对齐,其它与align_dlib_cpp保持一致。
@param five_points: 一个shape 为10 的人脸关键点坐标位置。
@attention 这个版本适合于两个眼角和鼻子，嘴角
'''
assert imgDim is not None
assert rgbImg is not None
from_points = []
for i in range(0,10,2):
from_points.append((five_points[i],five_points[i+1]))
to_points = []
for i in [36-17,45-17,33-17,48-17,54-17]:
new_ref_x = (self.padding+self.mean_shape_x[i])/(2*self.padding+1)
new_ref_y = (self.padding+self.mean_shape_y[i])/(2*self.padding+1)
to_points.append((imgDim *new_ref_x,imgDim *new_ref_y))
source = np.array(from_points).astype(np.int)
target = np.array(to_points,).astype(np.int)
source = np.reshape(source,(1,5,2))
target = np.reshape(target,(1,5,2))
H = cv2.estimateRigidTransform(source,target,False)
if H is None:
return
else:
aligned_face = cv2.warpAffine(rgbImg, H, (imgDim, imgDim))
return aligned_face
def align_similarity(self, imgDim, rgbImg, bb=None,
landmarks=None, landmarkIndices=range(0,68)):
'''
@brief 与align_openface()类似，但是通过更多个的点来计算仿射变换矩阵。
'''
assert imgDim is not None
assert rgbImg is not None
assert landmarkIndices is not None
if bb is None:
bb = self.getLargestFaceBoundingBox(rgbImg)
if bb is None:
return
if landmarks is None:
landmarks = self.findLandmarks(rgbImg, bb)
npLandmarkIndices = np.array(landmarkIndices)
npLandmarks = np.array(landmarks).astype(np.int)
T = (imgDim * self.new_template).astype(np.int)
source = np.reshape(npLandmarks[npLandmarkIndices],(1,68,2))
target = np.reshape(T[npLandmarkIndices],(1,68,2))
H = cv2.estimateRigidTransform(source,target,False)
if H is None:
return None
else:
aligned_face = cv2.warpAffine(rgbImg, H, (imgDim, imgDim))
return aligned_face
def getTransformationMatrixFromArray(A,B): # A,B both have to be numpy array # they must contain same number of points say n # so shape of A,B = [1,n,2] # returns a 3*3 np.matrix and a success flag(=1->OH yeah , 0 ->:()) T = cv2.estimateRigidTransform(A,B,False) return twoD_to_ThreeD_Affine(T)
def icp(d1, d2, max_iterate = 100):
src = np.array([d1.T], copy=True).astype(np.float32)
dst = np.array([d2.T], copy=True).astype(np.float32)
knn = cv2.KNearest()
responses = np.array(range(len(d2[0]))).astype(np.float32)
knn.train(src[0], responses)
Tr = np.array([[np.cos(0), -np.sin(0), 0],
[np.sin(0), np.cos(0), 0],
[0, 0, 1]])
dst = cv2.transform(dst, Tr[0:2])
max_dist = sys.maxint
scale_x = np.max(d1[0]) - np.min(d1[0])
scale_y = np.max(d1[1]) - np.min(d1[1])
scale = max(scale_x, scale_y)
for i in range(max_iterate):
ret, results, neighbours, dist = knn.find_nearest(dst[0], 1)
indeces = results.astype(np.int32).T
indeces = del_miss(indeces, dist, max_dist)
T = cv2.estimateRigidTransform(dst[0, indeces], src[0, indeces], True)
max_dist = np.max(dist)
dst = cv2.transform(dst, T)
Tr = np.dot(np.vstack((T,[0,0,1])), Tr)
if (is_converge(T, scale)):
break
return Tr[0:2]
def transform_marks(o_centers, new_centres, markers, shape, size=0.5):
bounds = getArea(new_centres, shape)
if new_centres is None: return markers
o_centers = centre_to_array(scale_centers(o_centers, size))
new_centres = centre_to_array(scale_centers(new_centres, size))
trs = cv2.estimateRigidTransform(o_centers, new_centres,True)
if trs is None: return markers
new_markers = np.zeros(markers.shape, dtype='uint8')
#downsize for computation
small_new = cv2.resize(new_markers, (0,0),fx=size, fy=size )
small_markers = cv2.resize(markers, (0,0),fx=size, fy=size)
for x in range(int(bounds[2]*size),int(bounds[3]*size)):
for y in range(int(bounds[0]*size),int(bounds[1]*size)):
try:
if small_markers[x][y] <= 0: continue
except: pass
c = np.array([[x],[y],[1]])
v = np.dot(trs,c)
if v[0]>=0 and v[0] < len(small_markers) and v[1]>=0 and v[1] < len(small_markers[0]):
small_new[int(v[0]),int(v[1])] = small_markers[x][y]
new_markers = cv2.resize(small_new, (0,0),fx=1/size, fy=1/size, interpolation = cv2.INTER_NEAREST)
return new_markers
def rigid_scaling_fom_points(fp, tp):
m = mean(fp[:2], axis=1)
maxstd = max(std(fp[:2], axis=1)) + 1e-9
C1 = diag([1 / maxstd, 1 / maxstd, 1])
C1[0][2] = -m[0] / maxstd
C1[1][2] = -m[1] / maxstd
fp = dot(C1, fp)
m = mean(tp[:2], axis=1)
maxstd = max(std(tp[:2], axis=1)) + 1e-9
C2 = diag([1 / maxstd, 1 / maxstd, 1])
C2[0][2] = -m[0] / maxstd
C2[1][2] = -m[1] / maxstd
tp = dot(C2, tp)
fp = float32(fp[:2,:]).T.reshape(1,-1)
tp = float32(tp[:2,:]).T.reshape(1,-1)
fp = fp.reshape([1,-1,2])
tp = tp.reshape([1,-1,2])
H = cv2.estimateRigidTransform(fp, tp, fullAffine=False)
H = vstack([H, [0, 0, 1]])
H = dot(linalg.inv(C2), dot(H, C1))
return H
def stitch_matrix(self, new_image):
'''stitch_matrix(new_image)
h -> homogeneous
i -> image
k -> keyframe
root -> root
'''
im_to_keyframe = cv2.estimateRigidTransform(new_image, self.keyframe_image, False)
if im_to_keyframe is None:
print '>Failed to map<'
return False, None
h_im_to_keyframe = self.make_homogeneous(im_to_keyframe)
ang_i2k, dx_i2k, dy_i2k, sx, sy = self.motion_from_matrix(im_to_keyframe)
if (np.fabs(sx - 1.0) > 0.01) or (np.fabs(sy - 1.0) > 0.01):
# print sx, sy
return False, h_im_to_keyframe
h_i2root = np.dot(self.h_k2root, h_im_to_keyframe)
ang_i2root, dx_i2root, dy_i2root, sxroot, syroot = self.motion_from_matrix(h_i2root)
rotated = self.rotate(new_image, np.degrees(ang_i2root), scale=sx)
# self.imshow("Rotated", rotated)
self.overlay(
rotated,
self.keyframe_position[0] + dy_i2root,
self.keyframe_position[1] + dx_i2root,
self.full_map
)
self.imshow("map", self.full_map)
return True, h_im_to_keyframe
def getPrevTransform(self):
ret, prev = self.cap.read()
prev_gray = cv2.cvtColor(prev, cv2.COLOR_BGR2GRAY)
p0 = cv2.goodFeaturesToTrack(prev_gray, mask = None, **feature_params)
while(True):
ret, cur = self.cap.read()
#print cur
if cur is None:
break
cur_gray = cv2.cvtColor(cur,cv2.COLOR_BGR2GRAY)
#print prev_corner
img0, img1 = prev_gray, cur_gray
# calculate optical flow
p1, st, err = cv2.calcOpticalFlowPyrLK(img0, img1, p0, None, **lk_params)
# Select good points
good_new = p1[st==1]
good_old = p0[st==1]
T = cv2.estimateRigidTransform(good_new, good_old, False)
print "---"
def getTransformationMatrixFromArray(A,B): # A,B both have to be numpy array # they must contain same number of points say n # so shape of A,B = [1,n,2] # returns a 3*3 np.matrix and a success flag(=1->OH yeah , 0 ->:()) T = cv2.estimateRigidTransform(A,B,False) #T2 = cv2.estimateRigidTransform(A,B,False) #T=(T1+T2)/2 dummy1,sizeA,dummy2 = A.shape dummy1,sizeB,dummy2 = B.shape X = np.empty([2,max(sizeA,sizeB),2]) X[0] = A X[1] = B # print X z =random.randint(0,10000) if(T is None): # waitForESC() textA = "FAILED_CHECK_A"+str(z) textB = "FAILED_CHECK_B"+str(z) #displayContour_transform(textA,A,700,700) #displayContour_transform(textB,B,700,700) return None textA = "SUCCESS_CHECK_A"+str(z) textB = "SUCCESS_CHECK_B"+str(z) ''' r1 = math.sqrt(T[0,0] * T[0,0] + T[0,1]*T[0,1]) T[0,0] =T[0,0]/r1 T[0,1] =T[0,1]/r1 r2 = math.sqrt(T[1,0] * T[1,0] + T[1,1]*T[1,1]) T[1,0] =T[1,0]/r2 T[1,1] =T[1,1]/r2 ''' #displayContour_transform(textA,A,700,700) #displayContour_transform(textB,B,700,700) return twoD_to_ThreeD_Affine(T)
def RANSAC_2D(point_set1, point_set2, iteration, tolerance):
point_count = point_set1.shape[0]
best_M = None
best_inlier = 0
best_model = None
for i in range(0, iteration):
sample_index = np.random.choice(range(0, point_count), 3, replace = False)
first_set = np.array([point_set1[sample_index]], dtype = np.float32)
second_set = np.array([point_set2[sample_index]], dtype = np.float32)
#print first_set.shape
transformation = cv2.estimateRigidTransform(first_set, second_set, fullAffine = False)
if transformation is None:
continue
print '******'
print first_set
print second_set
print '******'
#transformation[1,2] *= -1
new_point_set1 = np.asarray([np.dot(transformation, np.transpose(np.array([x[0], x[1], 1], dtype=np.float32))) for x in point_set1], dtype=np.float32)
distance_matrix = np.linalg.norm(new_point_set1 - point_set2, axis = 1)
inlier = np.sum(distance_matrix <= tolerance)
#print inlier
if inlier > best_inlier:
best_inlier = inlier
best_M = transformation
best_model = new_point_set1
print best_inlier
return best_M, best_model
def track_using_trajectories( cur, prev ):
global curr_loc_
global static_features_img_
p0 = cv2.goodFeaturesToTrack( cur, 200, 0.01, 5 )
insert_int_corners( p0 )
draw_point( cur, p0, 1 )
ellipse, p1 = update_mouse_location( p0 )
if p1 is not None:
for p in p1:
cv2.circle( cur, p, 10, 20, 2 )
cv2.ellipse( cur, ellipse, 1 )
cv2.circle( cur, curr_loc_, 10, 255, 3)
display_frame( cur, 1 )
# cv2.imshow( 'static features', static_features_img_ )
return
# Find a contour
prevE = find_edges( prev )
curE = find_edges( cur )
img = curE - prevE
cnts, hier = cv2.findContours( img, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE )
cnts = filter( ismouse, cnts )
cv2.drawContours( img, cnts, -1, 255, 3 )
display_frame( img, 1)
return
p1, status, err = cv2.calcOpticalFlowPyrLK( prev, cur, p0 )
mat = cv2.estimateRigidTransform( p0, p1, False )
# print cv2.warpAffine( curr_loc_, mat, dsize=(2,1) )
if mat is not None:
dx, dy = mat[:,2]
da = math.atan2( mat[1,0], mat[0,0] )
trajectory_.append( (dx, dy, da) )
print( "Transformation", dx, dy, da )
curr_loc_ = (curr_loc_[0] - int(dy), curr_loc_[1] - int(dx))
def getNorth(self, img_mat):
''' This method tries to estimate an
angle for the north, given a sample
image (img_mat, OpenCV matrix format).
The angle is given in radians, with
positive angles meaning clockwise
rotation of the north (counter clock-
wise rotation of the robot).
'''
kp2, des2 = self.sift.detectAndCompute(img_mat, None)
results = []
for kp1, des1 in self.ref_sifts:
matches = self.flann.knnMatch(des1, des2, k=2)
good = []
for m, n in matches:
if m.distance < 0.7*n.distance:
good.append(m)
results.append((kp1, des1, good))
kp1, des1, good = max(results, key=lambda x: x[2])
if len(good) > MIN_MATCHES:
src_pts = np.float32([ kp1[m.queryIdx].pt for m in good ]).reshape(-1,1,2)
dst_pts = np.float32([ kp2[m.trainIdx].pt for m in good ]).reshape(-1,1,2)
M = cv2.estimateRigidTransform(src_pts, dst_pts, False)
angle = np.arctan2(M[1,0],M[0,0])
return angle
else:
return None
def imRigidTransform(img, srcPts, dstPts):
srcPts = np.array([srcPts], np.int)
dstPts = np.array([dstPts], np.int)
M = cv2.estimateRigidTransform(srcPts, dstPts, False)
if transformation is not None:
return cv2.warpAffine(img, M)
else:
return None
def process(self, frame):
if self.prev is not None:
affine= cv2.estimateRigidTransform(frame, self.prev, True)
self.paff = _compose(affine, self.paff)
warp = cv2.warpAffine(frame, _translate(self.paff,(self.w/2)-60,(self.h/2)-48), (self.w, self.h))
self.out[warp>0] = warp[warp>0]
self.prev = frame
def stabilize (img_list):
cv_images = []
for i in xrange(len(img_list)):
cv_images.append(pil_to_opencv(img_list[i]))
#prev and prev_gray should be of type Mat
prev = cv_images[0]
cv2.cvtColor(prev, prev_gray, cv2.COLOR_BGR2GRAY)
for i in xrange(1, len(cv_images)):
current = cv_images[i]
cv2.cvtColor(current, current_gray, cv2.COLOR_BGR2GRAY)
cv2.goodFeaturesToTrack(prev_gray, prev_corner, 200, 0.01, 30)
cv2.calcOpticalFlowPyrLK(prev_gray, current_gray, prev_corner, current_corner, status, err)
prev_corner2 = []
current_corner2 = []
for j in xrange(len(status)):
if(status[j]):
prev_corner2.push_back(prev_corner[j])
current_corner2.push_back(current_corner[j])
T = cv2.estimateRigidTransform(prev_corner2, current_corner2, false)
if T.data == NULL:
last_T.copyTo(T)
T.copyTo(last_T)
dx = T.at(0,2)
dy = T.at(1,2)
da = numpy.atan2(T.at(1,0), T.at(0,0))
transform = []
transform.push_back(Transform(dx, dy, da))
current.copyTo(prev)
current_gray.copyTo(prev_gray)
# accumulate transformations to get image trajectory
x = 0
y = 0
a = 0
trajectory_list = []
for k in xrange(len(transform)):
x += transform[i].dx
y += transform[i].dy
a += transform[i].da
#trajectory_list.push_back(Trajectory(x,y,a))
return img_list
def findAffine(src, dst, fullAffine=False):
#print "src = %s" % str(src)
#print "dst = %s" % str(dst)
if len(src) >= affine_minpts:
affine = cv2.estimateRigidTransform(np.array([src]), np.array([dst]),
fullAffine)
else:
affine = None
#print str(affine)
return affine
def find_homography(four_pairs):
src, dst = split_src_dst(four_pairs)
src = np.float32(np.array(src)).reshape(-1, 1, 2)
dst = np.float32(np.array(dst)).reshape(-1, 1, 2)
transform = cv2.estimateRigidTransform(src, dst, False)
if transform is None:
return None
transform = np.vstack((transform, np.array([0, 0, 1])))
print 'rigid transform:', transform
return transform
def reestimate_transform(transform, points):
src, dst = split_src_dst(points)
src = np.float32(np.array(src)).reshape(-1, 1, 2)
dst = np.float32(np.array(dst)).reshape(-1, 1, 2)
transform = cv2.estimateRigidTransform(src, dst, False)
if transform is None:
return None
transform = np.vstack((transform, np.array([0, 0, 1])))
return transform
# Define objective as function of all points above and current transform (params)
'''
def getTransformParams( videoCapture, start=0, end=None ):
assert(videoCapture.isOpened())
# ensure that we start at the specified frame
videoCapture.set(cv2.cv.CV_CAP_PROP_POS_FRAMES, start)
maximum = int(videoCapture.get(cv2.cv.CV_CAP_PROP_FRAME_COUNT))
if end is None or end > maximum:
end = maximum
assert(end > start)
curr, currGray = None, None
prev, prevGray = None, None
_, prev = videoCapture.read()
prevGray = cv2.cvtColor(prev, cv2.COLOR_BGR2GRAY)
transforms = []
last = None
for i in range(end - start):
if not videoCapture.grab(): break
_, curr = videoCapture.retrieve()
currGray = cv2.cvtColor(curr, cv2.COLOR_BGR2GRAY)
prevPt, currPt = [], []
prevPtFilter, currPtFilter = [], []
prevPt = cv2.goodFeaturesToTrack(prevGray, mask=None, **FEATURE_PARAMS)
currPt, stats, _ = cv2.calcOpticalFlowPyrLK(prevGray, currGray, prevPt, None, **LK_PARAMS)
# get rid of bad matches
for i, stat in enumerate(stats):
if stat:
prevPtFilter.append(prevPt[i])
currPtFilter.append(currPt[i])
# estimate translation and rotation only
m = cv2.estimateRigidTransform(np.array(prevPtFilter), np.array(currPtFilter), False)
# check for transform not found
if m is None:
m = last.copy()
last = m.copy()
dx = m[0][2]
dy = m[1][2]
da = np.arctan2(m[1][0], m[0][0])
transforms.append((dx, dy, da))
prev = curr.copy()
prevGray = currGray.copy()
return transforms
def transform(refkp, framekp):
#add the extra [] to make 3D array
if refkp is None: return None
if framekp is None: return None
#if less than 5 matches then we can't find a transform
if len(refkp) < 5: return None
refkp = np.array([kp_to_xy_for_transform(refkp)],dtype = 'float32')
framekp = np.array([kp_to_xy_for_transform(framekp)], dtype = 'float32')
trs = cv2.estimateRigidTransform(refkp,framekp,False)
return trs
def getTransformationMatrixFrom_3_pairs(A,B,startA,middleA,endA,startB,middleB,endB): A3 = np.empty([1,3,2],np.int) B3 = np.empty([1,3,2],np.int) A3[0][1] = A.points[0][1] A3[0][2] = A.points[0][2] A3[0][3] = A.points[0][3] B3[0][1] = B.points[0][1] B3[0][2] = B.points[0][2] B3[0][3] = B.points[0][3] return cv2.estimateRigidTransform(A3,B3,False)
def _calculate_transform_matrix(self, other_hand):
src_points = self.getCentersList()
dest_points = other_hand.getCentersList()
min_points = min(len(src_points), len(dest_points))
src_points = src_points[:min_points]
dest_points = dest_points[:min_points]
src = np.float32(src_points).reshape(-1, 1, 2)
dest = np.float32(dest_points).reshape(-1, 1, 2)
transform = np.eye(3)
transform[:2, :] = cv2.estimateRigidTransform(src, dest, False)
return transform
def main(): shape = (1, 10, 2) # Needs to be a 3D array source = np.random.randint(0, 100, shape).astype(np.int) target = source + np.array([1, 0]).astype(np.int) print source print source.shape print source.dtype print source.itemsize print source.ndim print target.shape transformation = cv2.estimateRigidTransform(source, target, False) print transformation return
def getTransformationMatrixFromFragment_3pairs(A,B,startA,endA,startB,endB):
T = None
A3 = np.empty([1,3,2],np.int)
B3 = np.empty([1,3,2],np.int)
# print A.shape," ",startA," ",endA
A3[0][0] = A[startA]
A3[0][2] = A[endA]
B3[0][0] = B[startB]
B3[0][2] = B[endB]
middleA = (startA + endA)/2
middleB = (startB + endB)/2
A3[0][1] = A[middleA]
B3[0][1] = B[middleB]
T = cv2.estimateRigidTransform(A3,B3,False)
#print T
if(T is not None):
r1 = math.sqrt(T[0,0] * T[0,0] + T[0,1]*T[0,1])
T[0,0] =T[0,0]/r1
T[0,1] =T[0,1]/r1
r2 = math.sqrt(T[1,0] * T[1,0] + T[1,1]*T[1,1])
T[1,0] =T[1,0]/r2
T[1,1] =T[1,1]/r2
return twoD_to_ThreeD_Affine(T)
print "WE are in shit !"
print "*********************"
# print A3,B3
# def displayContour(imgname,contour,height,width):
img = np.zeros((800,800,1),np.uint8)
contour = np.empty([2,3,2],np.int)
contour[0] = A3
contour[1] = B3
# cv2.drawContours(img,contour,-1,255,1)
# cv2.imshow("Contours",img)
# k=cv2.waitKey(0)
# if(k==27):
# cv2.destroyAllWindows()
# return
print "*********************"
# i = (endA-middleA)
# while(T is None):
# return twoD_to_ThreeD_Affine(T)
return None
def estimateRigidTransform(srcPts, dstPts, inlierThreshold, outlierCoordScale = 1.0, fullAffine = False):
srcPts = np.float32(srcPts).reshape(-1,1,2)
dstPts = np.float32(dstPts).reshape(-1,1,2)
M = cv2.estimateRigidTransform(srcPts, dstPts, fullAffine = fullAffine)
print M
if M is None:
inlierMask = np.zeros(len(srcPts))
else:
inlierMask = []
mappedPts = cv2.transform(srcPts, M)
for mappedPt,dstPt in zip(mappedPts, dstPts):
dist = np.linalg.norm(mappedPt/outlierCoordScale - dstPt/outlierCoordScale)
inlierMask.append(int(dist < inlierThreshold))
inlierMask = np.array(inlierMask)
return M, inlierMask
def findAffine(self, i1, i2, pairs, fullAffine=False):
src = []
dst = []
for pair in pairs:
c1 = i1.coord_list[pair[0]]
c2 = i2.coord_list[pair[1]]
src.append( c1 )
dst.append( c2 )
#print "src = %s" % str(src)
#print "dst = %s" % str(dst)
affine = cv2.estimateRigidTransform(np.array([src]).astype(np.float32),
np.array([dst]).astype(np.float32),
fullAffine)
#print str(affine)
return affine
def estimate_partial_transform(matched_keypoints):
"""Wrapper of cv2.estimateRigidTransform for convenience in vidstab process
:param matched_keypoints: output of match_keypoints util function; tuple of (cur_matched_kp, prev_matched_kp)
:return: transform as list of [dx, dy, da]
"""
cur_matched_kp, prev_matched_kp = matched_keypoints
transform = cv2.estimateRigidTransform(np.array(prev_matched_kp),
np.array(cur_matched_kp), False)
if transform is not None:
# translation x
dx = transform[0, 2]
# translation y
dy = transform[1, 2]
# rotation
da = np.arctan2(transform[1, 0], transform[0, 0])
else:
dx = dy = da = 0
return [dx, dy, da]
def similarityTransform(inPoints, outPoints):
s60 = math.sin(60 * math.pi / 180)
c60 = math.cos(60 * math.pi / 180)
inPts = np.copy(inPoints).tolist()
outPts = np.copy(outPoints).tolist()
xin = c60 * (inPts[0][0] - inPts[1][0]) - s60 * (inPts[0][1] -
inPts[1][1]) + inPts[1][0]
yin = s60 * (inPts[0][0] - inPts[1][0]) + c60 * (inPts[0][1] -
inPts[1][1]) + inPts[1][1]
inPts.append([np.int(xin), np.int(yin)])
xout = c60 * (outPts[0][0] - outPts[1][0]) - s60 * (
outPts[0][1] - outPts[1][1]) + outPts[1][0]
yout = s60 * (outPts[0][0] - outPts[1][0]) + c60 * (
outPts[0][1] - outPts[1][1]) + outPts[1][1]
outPts.append([np.int(xout), np.int(yout)])
tform = cv2.estimateRigidTransform(np.array([inPts]), np.array([outPts]),
False)
return tform
def similarityTransform(inPoints, outPoints): s60 = math.sin(60*math.pi/180) c60 = math.cos(60*math.pi/180) inPts = np.copy(inPoints).tolist() outPts = np.copy(outPoints).tolist() # The third point is calculated so that the three points make an equilateral triangle xin = c60*(inPts[0][0] - inPts[1][0]) - s60*(inPts[0][1] - inPts[1][1]) + inPts[1][0] yin = s60*(inPts[0][0] - inPts[1][0]) + c60*(inPts[0][1] - inPts[1][1]) + inPts[1][1] inPts.append([np.int(xin), np.int(yin)]) xout = c60*(outPts[0][0] - outPts[1][0]) - s60*(outPts[0][1] - outPts[1][1]) + outPts[1][0] yout = s60*(outPts[0][0] - outPts[1][0]) + c60*(outPts[0][1] - outPts[1][1]) + outPts[1][1] outPts.append([np.int(xout), np.int(yout)]) # Now we can use estimateRigidTransform for calculating the similarity transform. tform = cv2.estimateRigidTransform(np.array([inPts]), np.array([outPts]), False) return tform
def compute_trajectory(filename):
if 'mp4' in filename:
pkl = filename.replace('mp4', 'pkl')
else:
pkl = filename.replace('avi', 'pkl')
if os.path.exists(pkl):
transform = pickle.load(open(pkl, 'rb'))
return transform
cap = cv2.VideoCapture(filename)
ret, frame_prev = cap.read()
frame_prev = cv2.cvtColor(frame_prev, cv2.COLOR_BGR2GRAY)
point_prev = cv2.goodFeaturesToTrack(frame_prev,
mask=None,
**feature_params)
transform = []
while True:
ret, frame = cap.read()
if not ret:
break
frame_cur = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
point_cur, status, err = cv2.calcOpticalFlowPyrLK(
frame_prev, frame_cur, point_prev, None, **lk_params)
good_cur = point_cur[status == 1]
good_prev = point_prev[status == 1]
trans = cv2.estimateRigidTransform(good_prev, good_cur, False)
if trans is None:
transform.append(transform[-1])
else:
transform.append(trans)
frame_prev = frame_cur
point_prev = cv2.goodFeaturesToTrack(frame_prev,
mask=None,
**feature_params)
cap.release()
transform = np.array(transform)
pickle.dump(transform, open(pkl, 'wb'))
return transform
def align_dlib_cpp(self, rgbImg, landmarks=None):
'''
@brief: 与dlib C++版本实现的裁剪对齐方法一致。
@attention
'''
assert rgbImg is not None
npLandmarks = np.array(landmarks)[:, :2]
shape_x = [npLandmarks[i][0] for i in range(68)]
shape_y = [npLandmarks[i][1] for i in range(68)]
from_points = []
to_points = []
for i in range(17, 68):
# 忽略掉低于嘴唇的部分
if i >= 55 and i <= 59:
continue
# 忽略眉毛部分
if i >= 17 and i <= 26:
continue
# 上下左右都padding
new_ref_x = (self.padding +
self.mean_shape_x[i - 17]) / (2 * self.padding + 1)
new_ref_y = (self.padding +
self.mean_shape_y[i - 17]) / (2 * self.padding + 1)
from_points.append((shape_x[i], shape_y[i]))
to_points.append(
(self.image_size * new_ref_x, self.image_size * new_ref_y))
source = np.array(from_points).astype(np.int)
target = np.array(to_points, ).astype(np.int)
source = np.reshape(source, (1, 36, 2))
target = np.reshape(target, (1, 36, 2))
H = cv2.estimateRigidTransform(source, target, False)
if H is None:
return None
else:
aligned_face = cv2.warpAffine(rgbImg, H,
(self.image_size, self.image_size))
return aligned_face
def RANSAC_2D(point_set1, point_set2, iteration, tolerance):
point_count = point_set1.shape[0]
best_M = None
best_inlier = 0
best_model = None
for i in range(0, iteration):
sample_index = np.random.choice(range(0, point_count),
3,
replace=False)
first_set = np.array([point_set1[sample_index]], dtype=np.float32)
second_set = np.array([point_set2[sample_index]], dtype=np.float32)
#print first_set.shape
transformation = cv2.estimateRigidTransform(first_set,
second_set,
fullAffine=False)
if transformation is None:
continue
#transformation[0, 1] *= -1
#transformation[1, 0] *= -1
print '******'
print first_set
print second_set
print '******'
#transformation[1,2] *= -1
new_point_set1 = np.asarray([
np.dot(transformation,
np.transpose(np.array([x[0], x[1], 1], dtype=np.float32)))
for x in point_set1
],
dtype=np.float32)
distance_matrix = np.linalg.norm(new_point_set1 - point_set2, axis=1)
inlier = np.sum(distance_matrix <= tolerance)
#print inlier
if inlier > best_inlier:
best_inlier = inlier
best_M = transformation
best_model = new_point_set1
print best_inlier
return best_M, best_model
def align(image_stack):
"""
Align the image stack with rigid affine transforms between frames
"""
log.info("Aligning images with rigid transform.")
# calculate the image corners to auto-crop frames;
# only using the northwest and southeast corners can lead to
# errors if the other two corners form a smaller bounding box
(r, c) = image_stack[0].shape[:2]
corners = np.array([[0., c], [0., r], [1., 1.]], dtype=np.float)
nw_corner, se_corner = corners[:, 0], corners[:, 1]
transform = np.eye(3)
_stack = [image_stack[0]]
# iterate over each pair of images and compute the cumulative
# rigid transform to project the current frame onto the first
# image frame
for anchor, img in zip(image_stack[:-1], image_stack[1:]):
new_t = cv2.estimateRigidTransform(img, anchor, fullAffine=False)
transform[:2, :2] = new_t[:2, :2].dot(transform[:2, :2])
transform[:2, 2] = new_t[:, 2] + transform[:2, 2]
new_im = cv2.warpAffine(img, transform[:2, :], (c, r))
_stack.append(new_im)
bounds = transform.dot(corners)
nw_corner = np.max([nw_corner, bounds[:, 0]], axis=0)
se_corner = np.min([se_corner, bounds[:, 1]], axis=0)
log.debug("Transformation matrix:\n" + str(transform))
lx, ly = nw_corner[:2]
rx, ry = se_corner[:2]
_stack = [img[ly:ry, lx:rx] for img in _stack]
return _stack, _stack[0].shape
def callback(image):
bridge = CvBridge()
global first
if first:
print("firsted")
global first_image
first_img = bridge.imgmsg_to_cv2(image, "bgr8")
global prev
prev = first_img
global first
ref_set.append({'img': image, 'tf': [0,0,0,0,0]})
first = False
else:
transformation = Transform()
global first_img
global prev
curr = bridge.imgmsg_to_cv2(image, "bgr8")
#new = first_img[:, 60:300]
test1 = deepcopy(first_img)
test2 = deepcopy(first_img)
test1 = test1[30:209, 40:279]
test2 = test2[35:214, 40:279]
affine = cv2.estimateRigidTransform(test1, test2, False)
print(affine)
if affine is not None:
transform = affineToTransform(affine, summed_transformation)
# Summing
summed_transformation.translation.x += transform.translation.x
summed_transformation.translation.y *= transform.translation.y
summed_transformation.translation.z += transform.translation.z
summed_transformation.rotation.x += transform.rotation.x
summed_transformation.rotation.y += transform.rotation.y
summed_transformation.rotation.z += transform.rotation.z
summed_transformation.rotation.w += transform.rotation.w
#print(summed_transformation)
cmdpub.publish(summed_transformation)
global prev
prev = curr
else:
print("skipped")
def compute_location(self, kp1, des1, kp2, des2):
"""
compute the global location of center of current image
:param kp1: captured keyPoints
:param des1: captured descriptions
:param kp2: map keyPoints
:param des2: map descriptions
:return: global pose
"""
good = []
pose = None
if des1 is not None and des2 is not None:
matches = self.matcher.knnMatch(des1, des2, k=2)
for match in matches:
if len(match) > 1 and match[0].distance < MATCH_RATIO * match[1].distance:
good.append(match[0])
if len(good) > MIN_MATCH_COUNT:
src_pts = np.float32([kp1[m.queryIdx].pt for m in good]).reshape(-1, 1, 2)
dst_pts = np.float32([kp2[m.trainIdx].pt for m in good]).reshape(-1, 1, 2)
transform = cv2.estimateRigidTransform(src_pts, dst_pts, False)
if transform is not None:
transformed_center = cv2.transform(CAMERA_CENTER, transform) # get global pixel
transformed_center = [transformed_center[0][0][0] / METER_TO_PIXEL, # map to global pose
(MAP_PIXEL_HEIGHT - 1 - transformed_center[0][0][1]) / METER_TO_PIXEL]
yaw = np.arctan2(transform[1, 0], transform[0, 0]) # get global heading
# correct the pose if the drone is not level
z = math.sqrt(self.z ** 2 / (1 + math.tan(self.angle_x) ** 2 + math.tan(self.angle_y) ** 2))
offset_x = np.tan(self.angle_x) * z
offset_y = np.tan(self.angle_y) * z
global_offset_x = math.cos(yaw) * offset_x + math.sin(yaw) * offset_y
global_offset_y = math.sin(yaw) * offset_x + math.cos(yaw) * offset_y
pose = [transformed_center[0] + global_offset_x, transformed_center[1] + global_offset_y, z, yaw]
return pose, len(good)
def findMinDistFace(landmarks1, landmarks2):
faceind = np.zeros((len(landmarks2), ))
for i in range(len(landmarks2)):
if landmarks2[i] is None:
continue
mindist = np.inf
for j in range(len(landmarks1)):
if landmarks1[j] is None:
continue
T = cv2.estimateRigidTransform(landmarks1[j], landmarks2[i], False)
if T is None:
continue
T_full = np.vstack((T, np.array([0, 0, 1])))
landmarks1_full = np.vstack(
(landmarks1[j].T, np.ones((1, landmarks1[j].shape[0]))))
landmarks1_trans = np.dot(T_full, landmarks1_full)
landmarks1_trans = landmarks1_trans[0:2, :].T
dist = calcSimilarity(landmarks1_trans, landmarks2[i])
if dist < mindist:
faceind[i] = j
mindist = dist
return faceind.astype(int)
def get_cov_from_video(video_name, size):
cap, n_frames, fps, prev = video_open(video_name, size)
new_size = size
old = []
last_affine = ...
cumulative_transform = np.insert(np.array([[1, 0], [0, 1]]), [2], [0], axis=1)
for i in range(n_frames-1):
# read frames
ret2, cur = cap.read()
cur = cv2.resize(cur, new_size, cv2.INTER_CUBIC)
# get affine transform between frames
affine = cv2.estimateRigidTransform(prev, cur, False)
# Sometimes there is no Affine transform between frames, so we use the last
if not np.all(affine):
affine = last_affine
last_affine = affine
# Accumulated frame to frame original transform
#cumulative_transform = sum_2_affine(cumulative_transform, affine)
# save original affine for comparing with stabilized
old.append(affine)
cov = covariance(*get_params_from_trajectory(old))
return cov
def tran_similarity(src_img, src_points, dst_img, dst_points): # 获取人脸平均脸
s60 = math.sin(60 * math.pi / 180)
c60 = math.cos(60 * math.pi / 180)
in_pts = np.copy(dst_points).tolist()
out_pts = np.copy(src_points).tolist()
x_in = c60 * (in_pts[0][0] - in_pts[1][0]) - s60 * (in_pts[0][1] - in_pts[1][1]) + in_pts[1][0]
y_in = s60 * (in_pts[0][0] - in_pts[1][0]) + c60 * (in_pts[0][1] - in_pts[1][1]) + in_pts[1][1]
in_pts.append([np.int(x_in), np.int(y_in)])
x_out = c60 * (out_pts[0][0] - out_pts[1][0]) - s60 * (out_pts[0][1] - out_pts[1][1]) + out_pts[1][0]
y_out = s60 * (out_pts[0][0] - out_pts[1][0]) + c60 * (out_pts[0][1] - out_pts[1][1]) + out_pts[1][1]
out_pts.append([np.int(x_out), np.int(y_out)])
m = cv2.estimateRigidTransform(np.array([in_pts]), np.array([out_pts]), False)
output = cv2.warpAffine(dst_img, m, (src_img.shape[1], src_img.shape[0]))
return output
def compute_transform(matcher, kp1, des1, kp2, des2):
"""
computes the transformation between two sets of keypoints and descriptors
"""
transform = None
if des1 is not None and des2 is not None:
matches = matcher.knnMatch(des1, des2, k=2)
good = []
for match in matches:
if len(match) > 1 and match[0].distance < MATCH_RATIO * match[1].distance:
good.append(match[0])
src_pts = np.float32([kp1[m.queryIdx] for m in good]).reshape(-1, 1, 2)
dst_pts = np.float32([kp2[m.trainIdx] for m in good]).reshape(-1, 1, 2)
# estimateRigidTransform needs at least three pairs
if src_pts is not None and dst_pts is not None and len(src_pts) > 3 and len(dst_pts) > 3:
transform = cv2.estimateRigidTransform(src_pts, dst_pts, False)
return transform
def align(imgDim, rgbImg, detectedLandmarks, templateLandmarks):
assert imgDim is not None
assert rgbImg is not None
assert templateLandmarks is not None
detectedLandmarks = np.float32(np.array(detectedLandmarks))
templateLandmarks = np.float32(np.int32(np.array(templateLandmarks)))
# if landmarks are not detected, use previous frames' warp matrix
if (len(detectedLandmarks) > 0):
H = cv2.estimateRigidTransform(detectedLandmarks, templateLandmarks,
True)
else:
H = None
SMOOTH_NUM = 3
if (len(pastH) < SMOOTH_NUM):
if H is None:
return []
for ii in range(SMOOTH_NUM):
pastH.append(H)
smoothH = np.zeros((2, 3))
if H is not None:
pastH.pop()
pastH.insert(0, H)
# Take average of last 'SMOOTH_NUM' warp matrices to reduce jitter
for ii in range(SMOOTH_NUM):
smoothH = smoothH + pastH[ii]
smoothH = smoothH / SMOOTH_NUM
transformed = cv2.warpAffine(rgbImg, smoothH, (imgDim, imgDim))
return transformed
def find_transformation(previous_points, current_points):
"""finds the transformation matrix from previous points to current points.
The transformation matrix is found using estimateRigidTransform
(fancier alternatives have been tried, but are not that stable).
Parameters
----------
previous_points: numpy.ndarray
Set of 'starting' 2d points
current_points: numpy.ndarray
Set of 'destination' 2d points
Returns
-------
transformation_matrix: numpy.ndarray
the affine transformation matrix between
the two sets of points.
"""
from cv2 import estimateRigidTransform
return estimateRigidTransform(previous_points, current_points, False)
def findGroupAffine(self, i1, fullAffine=False):
# find the affine transform matrix representing the best fit
# against all the placed neighbors. Builds a cumulative
# src/dest list with our src points listed once for each image
# pair.
src = []
dst = []
for i, pairs in enumerate(i1.match_list):
if len(pairs) < 3:
# can't compute affine transform on < 3 points
continue
i2 = self.image_list[i]
if not i2.placed:
# don't consider non-yet-placed neighbors
continue
# add coordinate matches for this image pair
for pair in pairs:
c1 = i1.coord_list[pair[0]]
c2 = i2.coord_list[pair[1]]
src.append( c1 )
dst.append( c2 )
if len(src) < 3:
# not enough points to compute affine transformation
return np.array( [ [1.0, 0.0, 0.0 ], [0.0, 1.0, 0.0] ] )
# find the affine matrix on the communlative set of all
# matching coordinates for all matching image pairs
# simultaneously...
affine = cv2.estimateRigidTransform(np.array([src]).astype(np.float32),
np.array([dst]).astype(np.float32),
fullAffine)
if affine == None:
# it's possible given a degenerate point set, the affine
# estimator will return None, so return the identity
affine = np.array( [ [1.0, 0.0, 0.0 ], [0.0, 1.0, 0.0] ] )
return affine
def calculateAffinityMatrixAndDraw(bestImage, inliersDataBase, inliersWebCam,
imgout):
# The affinity mat A
A = cv2.estimateRigidTransform(inliersDataBase,
inliersWebCam,
fullAffine=True)
A = np.vstack((A, [0, 0, 1]))
# Calculate the points of the rectangle occupied by the recognized object
a = np.array([0, 0, 1], np.float)
b = np.array([bestImage.shape[1], 0, 1], np.float)
c = np.array([bestImage.shape[1], bestImage.shape[0], 1], np.float)
d = np.array([0, bestImage.shape[0], 1], np.float)
center = np.array(
[float(bestImage.shape[0]) / 2,
float(bestImage.shape[1]) / 2, 1], np.float)
# Multiply the points of the virtual space, to convert them into
# real image points
a = np.dot(A, a)
b = np.dot(A, b)
c = np.dot(A, c)
d = np.dot(A, d)
center = np.dot(A, center)
# The points are dehomogenized
areal = (int(a[0] / a[2]), int(a[1] / b[2]))
breal = (int(b[0] / b[2]), int(b[1] / b[2]))
creal = (int(c[0] / c[2]), int(c[1] / c[2]))
dreal = (int(d[0] / d[2]), int(d[1] / d[2]))
centerreal = (int(center[0] / center[2]), int(center[1] / center[2]))
# The polygon and the file name of the image are painted in the middle of the polygon
points = np.array([areal, breal, creal, dreal], np.int32)
cv2.polylines(imgout, np.int32([points]), 1, (0, 255, 255), thickness=3)
# Obj_utilscv.draw_str(imgout, centerreal, bestImage.nameFile.upper())
# The detected object is displayed in a separate window
cv2.imshow('ImageDetector', bestImage.imageBinary)
def getdrift2(green, sample_interval=100):
''' This works by creating a list of frames averaged by every sample_interval.
This is a more robust solution than getdrift1'''
frames = []
nframes = int(np.ceil(green.shape[0] / sample_interval) + 1)
#create the list of frames
for n in np.arange(nframes):
f = green[n * sample_interval:(n + 1) * sample_interval].mean(0)
frames.append((f * 256 / green.max()).astype('uint8'))
allvectors = []
for i in range(nframes - 1):
motionVectors = []
for j in range(nframes - 1):
flow = cv2.estimateRigidTransform(frames[i], frames[j], False)
motionVectors.append([flow[0, 2], flow[1, 2]])
motionVectors = np.array(motionVectors)
allvectors.append(motionVectors)
# take the derivative so all the vectors look similar
allvectors_d = []
for v in allvectors:
motionVectors_d = [np.array([0, 0])]
for f in np.arange(nframes - 2) + 1:
motionVectors_d.append(v[f] - v[f - 1])
motionVectors_d = np.array(motionVectors_d)
allvectors_d.append(motionVectors_d)
allvectors_d = np.array(allvectors_d)
motionVectors = np.median(allvectors_d, 0)
# take the integral so we have position rather than velocity
for i in np.arange(motionVectors.shape[0] - 1) + 1:
motionVectors[i] = motionVectors[i] + motionVectors[
i -
1] #take the integral, so we have position rather than velocity
motionVectors = np.repeat(motionVectors, sample_interval, axis=0)
motionVectors2 = smoothMotion(motionVectors, 3) #fit with polynomial
return motionVectors2
def similarityTransform(inPoints, outPoints):
"""
求解相似矩阵
参数:
==========
inPoints:输入点
outPoints:输出点
返回值
==========
tform:相似矩阵
"""
pass
s60 = math.sin(60*math.pi/180)
c60 = math.cos(60*math.pi/180)
inPts= np.copy(inPoints).tolist()
outPts = np.copy(outPoints).tolist()
xin= c60*(inPts[1][0]-inPts[0][0])+s60*(inPts[1][1]-inPts[0][1])+inPts[0][0]
yin = c60*(inPts[1][1]-inPts[0][1])-s60*(inPts[1][0]-inPts[0][0])+inPts[0][1]
inPts.append([np.int(xin), np.int(yin)])
xout= c60*(outPts[1][0]-outPts[0][0])+s60*(outPts[1][1]-outPts[0][1])+outPts[0][0]
yout = c60*(outPts[1][1]-outPts[0][1])-s60*(outPts[1][0]-outPts[0][0])+outPts[0][1]
outPts.append([np.int(xout), np.int(yout)])
if cv2.__version__ > '3.5':
print('opencv 版本过高,使用estimateAffinePartial2D方法代替estimateRigidTransform')
_tform = cv2.estimateAffinePartial2D(np.array(inPts), np.array(outPts))
tform = _tform[0]
else:
tform = cv2.estimateRigidTransform(np.array(inPts), np.array(outPts), False)
# tform = cv2.estimateAffinePartial2D(np.array(inPts),np.array(outPts),False)
return tform
def calculateAffinityMatrixAndDraw(bestImage, inliersDataBase, inliersWebCam,
imgout):
#Se calcula la matriz de afinidad A
A = cv2.estimateRigidTransform(inliersDataBase,
inliersWebCam,
fullAffine=True)
A = np.vstack((A, [0, 0, 1]))
#Se calculan los puntos del rectangulo que ocupa el objeto reconocido
a = np.array([0, 0, 1], np.float)
b = np.array([bestImage.shape[1], 0, 1], np.float)
c = np.array([bestImage.shape[1], bestImage.shape[0], 1], np.float)
d = np.array([0, bestImage.shape[0], 1], np.float)
centro = np.array(
[float(bestImage.shape[0]) / 2,
float(bestImage.shape[1]) / 2, 1], np.float)
#Se multiplican los puntos del espacio virtual, para convertirlos en
#puntos reales de la imagen
a = np.dot(A, a)
b = np.dot(A, b)
c = np.dot(A, c)
d = np.dot(A, d)
centro = np.dot(A, centro)
#Se deshomogeneizan los puntos
areal = (int(a[0] / a[2]), int(a[1] / b[2]))
breal = (int(b[0] / b[2]), int(b[1] / b[2]))
creal = (int(c[0] / c[2]), int(c[1] / c[2]))
dreal = (int(d[0] / d[2]), int(d[1] / d[2]))
centroreal = (int(centro[0] / centro[2]), int(centro[1] / centro[2]))
#Se pinta el polígono y el nombre del fichero de la imagen en el centro del polígono
points = np.array([areal, breal, creal, dreal], np.int32)
cv2.polylines(imgout, np.int32([points]), 1, (255, 255, 255), thickness=2)
utilscv.draw_str(imgout, centroreal, bestImage.nameFile.upper())
#Se visualiza el objeto detectado en una ventana a parte
cv2.imshow('ImageDetector', bestImage.imageBinary)
def transform_image_from_points(img: np.array,
matches: list,
kp_ref: list,
kp_img: list,
transform: str = "rigid"):
"""
transfroms the image `img` using a list of keypoints from a source and target image.
uses the optional parameter "transform" which if set to "h**o", will use a homography transform other than a rigid
one.
:param img: image to transform
:type: np.array
:param matches: list of matches
:type: list(cv2.Match)
:param kp_ref: list of keypoints for the reference points
:param kp_img: list of keypoints for the target image points
:param transform: how to transform the image
:return: transformed image
:rtype: np.array
"""
src_pts = np.float32([kp_ref[m.queryIdx].pt
for m in matches]).reshape(-1, 1, 2)
dst_pts = np.float32([kp_img[m.trainIdx].pt
for m in matches]).reshape(-1, 1, 2)
if transform == "h**o":
# homography transform, better
M, mask = cv2.findHomography(src_pts, dst_pts, cv2.RANSAC, 20.0)
return cv2.warpPerspective(img,
M, (img.shape[1], img.shape[0]),
flags=cv2.INTER_CUBIC +
cv2.WARP_INVERSE_MAP)
else:
# rigid transform,
M = cv2.estimateRigidTransform(src_pts, dst_pts, False)
return cv2.warpAffine(img,
M, (img.shape[1], img.shape[0]),
flags=cv2.INTER_CUBIC + cv2.WARP_INVERSE_MAP)
def feature_transform(self, kp1, des1, kp2, des2):
transform = None
if des1 is not None and des2 is not None:
matches = self.matcher.knnMatch(des1, des2, k=2)
good = []
for match in matches:
if len(match
) > 1 and match[0].distance < 0.7 * match[1].distance:
good.append(match[0])
src_pts = np.float32([kp1[m.queryIdx].pt
for m in good]).reshape(-1, 1, 2)
dst_pts = np.float32([kp2[m.trainIdx].pt
for m in good]).reshape(-1, 1, 2)
# estimateRigidTransform needs at least three pairs
if src_pts is not None and dst_pts is not None and len(
src_pts) > 3 and len(dst_pts) > 3:
transform = cv2.estimateRigidTransform(src_pts, dst_pts, False)
return transform
def similarityTransform(inPoints, outPoints):
"""
求解相似矩阵
参数:
==========
inPoints:输入点
outPoints:输出点
返回值
==========
tform:相似矩阵
"""
pass
s60 = math.sin(60 * math.pi / 180)
c60 = math.cos(60 * math.pi / 180)
inPts = np.copy(inPoints).tolist()
outPts = np.copy(outPoints).tolist()
xin = c60 * (inPts[1][0] - inPts[0][0]) + s60 * (inPts[1][1] -
inPts[0][1]) + inPts[0][0]
yin = c60 * (inPts[1][1] - inPts[0][1]) - s60 * (inPts[1][0] -
inPts[0][0]) + inPts[0][1]
inPts.append([np.int(xin), np.int(yin)])
xout = c60 * (outPts[1][0] - outPts[0][0]) + s60 * (
outPts[1][1] - outPts[0][1]) + outPts[0][0]
yout = c60 * (outPts[1][1] - outPts[0][1]) - s60 * (
outPts[1][0] - outPts[0][0]) + outPts[0][1]
outPts.append([np.int(xout), np.int(yout)])
tform = cv2.estimateRigidTransform(np.array(inPts), np.array(outPts),
False)
return tform
def getRigidTransform(img1, img2):
# find the keypoints and descriptors with SIFT
sift = cv2.xfeatures2d.SIFT_create()
kp1, des1 = sift.detectAndCompute(img1, None)
kp2, des2 = sift.detectAndCompute(img2, None)
# FLANN parameters
FLANN_INDEX_KDTREE = 0
index_params = dict(algorithm=FLANN_INDEX_KDTREE, trees=5)
search_params = dict(checks=50)
flann = cv2.FlannBasedMatcher(index_params, search_params)
# Find nearest 2 neighbors
feature_matches = flann.knnMatch(des1, des2, k=2)
# store all the good matches as per Lowe's ratio test.
good = []
for m, n in feature_matches:
if m.distance < 0.7 * n.distance:
good.append(m)
img1_pts = np.float32([kp1[m.queryIdx].pt for m in good]).reshape(-1, 2)
img2_pts = np.float32([kp2[m.trainIdx].pt for m in good]).reshape(-1, 2)
# orb = cv2.ORB_create()
# kp1, des1 = orb.detectAndCompute(img1,None)
# kp2, des2 = orb.detectAndCompute(img2,None)
# bf = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=True)
# matches = bf.match(des1,des2)
# img1_pts = np.float32([ kp1[m.queryIdx].pt for m in matches ]).reshape(-1,2)
# img2_pts = np.float32([ kp2[m.trainIdx].pt for m in matches ]).reshape(-1,2)
A = cv2.estimateRigidTransform(img1_pts, img2_pts, False)
if A is None:
return 0, 0, 0
dx = A[0, 2]
dy = A[1, 2]
da = np.arctan2(A[1, 0], A[0, 0])
return dx, dy, da
def icp(d1, d2, max_iterate=100):
src = np.array([d1.T], copy=True).astype(np.float32)
dst = np.array([d2.T], copy=True).astype(np.float32)
#knn = cv2.KNearest()
knn = cv2.ml.KNearest_create()
# knn = NearestNeighbors()
responses = np.array(range(len(d2[0]))).astype(np.float32)
knn.train(src[0], responses)
Tr = np.array([[np.cos(0), -np.sin(0), 0], [np.sin(0),
np.cos(0), 0], [0, 0, 1]])
dst = cv2.transform(dst, Tr[0:2])
max_dist = sys.maxint
scale_x = np.max(d1[0]) - np.min(d1[0])
scale_y = np.max(d1[1]) - np.min(d1[1])
scale = max(scale_x, scale_y)
for i in range(max_iterate):
ret, results, neighbours, dist = knn.find_nearest(dst[0], 1)
indeces = results.astype(np.int32).T
indeces = del_miss(indeces, dist, max_dist)
T = cv2.estimateRigidTransform(dst[0, indeces], src[0, indeces], True)
max_dist = np.max(dist)
dst = cv2.transform(dst, T)
Tr = np.dot(np.vstack((T, [0, 0, 1])), Tr)
if (is_converge(T, scale)):
break
return Tr[0:2]
def ransac_rigid(source_points, target_points, confidence, threshold):
try:
max_iters = 25000
transform, inliers = cv2.estimateAffinePartial2D(source_points, target_points, 0, ransacReprojThreshold = threshold, maxIters = max_iters, confidence = confidence)
source_points = np.squeeze(source_points, None)
target_points = np.squeeze(target_points, None)
transform = cv2.estimateRigidTransform(
np.resize(source_points[matlib.repmat(inliers.astype(bool), 1, 2)],
(np.sum(inliers), 2)),
np.resize(target_points[matlib.repmat(inliers.astype(bool), 1, 2)],
(np.sum(inliers), 2)),
0)
if transform is not None:
t_transform = transform
transform = np.eye(3)
transform[0:2, 0:3] = t_transform
failed = False
else:
transform = np.eye(3)
failed = True
except:
transform = np.eye(3)
failed = True
return transform, failed
def face_align(_image, _landmark, _target_shape):
"""
人脸对齐
Args:
_image: 人脸图片
_landmark: 人脸图片上的landmark
_target_shape: 目标尺寸
Returns: 对齐后的人脸
"""
reference_facial_points = np.array(
[[0.31556875, 0.4615741], [0.6826229, 0.45983392],
[0.5002625, 0.6405054], [0.3494719, 0.82469195],
[0.6534365, 0.8232509]],
dtype=np.float32)
target_facial_points = reference_facial_points.copy() * _target_shape
h, w = _image.shape[:2]
remapped_landmark = _landmark.copy() * [w, h]
transform_matrix = cv2.estimateRigidTransform(remapped_landmark,
target_facial_points, True)
face_img = cv2.warpAffine(_image, transform_matrix, _target_shape)
return face_img
def icp(a, b, init_pose=(0, 0, 0), no_iterations=13):
src = np.array([a.T], copy=True).astype(np.float32)
dst = np.array([b.T], copy=True).astype(np.float32)
#Initialise with the initial pose estimation
Tr = np.array([[np.cos(init_pose[2]), -np.sin(init_pose[2]), init_pose[0]],
[np.sin(init_pose[2]),
np.cos(init_pose[2]), init_pose[1]], [0, 0, 1]])
src = cv2.transform(src, Tr[0:2])
for i in range(no_iterations):
#Find the nearest neighbours between the current source and the
#destination cloudpoint
nbrs = NearestNeighbors(n_neighbors=1, algorithm='auto').fit(dst[0])
distances, indices = nbrs.kneighbors(src[0])
#Compute the transformation between the current source
#and destination cloudpoint
T = cv2.estimateRigidTransform(src, dst[0, indices.T], False)
#Transform the previous source and update the
#current source cloudpoint
src = cv2.transform(src, T)
#Save the transformation from the actual source cloudpoint
#to the destination
Tr = np.dot(Tr, np.vstack((T, [0, 0, 1])))
return Tr[0:2]
def rigid_estimation(mov):
nbFrames = len(mov)
file = open("outputFiles/outTransform.txt", "w")
lk_params = dict(winSize = (15,15),
maxLevel = 2,
criteria = (cv2.TERM_CRITERIA_EPS | cv2.TERM_CRITERIA_COUNT, 10, 0.03))
# retrieve initial frame
prev = mov[0]
gray_prev = cv2.cvtColor(prev, cv2.COLOR_BGR2GRAY)
last_T = np.matrix('1 0 0; 0 1 0')
prev_to_cur_transform = []
# iterate over all frames
k = 1
for t in range(1, nbFrames):
# retrieve current frame
cur = mov[t]
gray_cur = cv2.cvtColor(cur, cv2.COLOR_BGR2GRAY)
# find features to track in previous frames
prev_corner = cv2.goodFeaturesToTrack(gray_prev, 200, 0.01, 30)
# calculate new coordinates of features in current frame
cur_corner, status, err = cv2.calcOpticalFlowPyrLK(gray_prev, gray_cur,
prev_corner, None,
**lk_params)
# remove outliers
prev_corner1 = []
cur_corner1 = []
for s in range(0, len(status)):
if status[s]:
prev_corner1.append(prev_corner[s])
cur_corner1.append(cur_corner[s])
print "Frame: " + str(k) + "/" + str(nbFrames) + " - good optical flow: " + str(len(prev_corner1));
# estimate affine transform between the frames
prev_corner1 = np.asarray(prev_corner1)
cur_corner1 = np.asarray(cur_corner1)
T = cv2.estimateRigidTransform(prev_corner1, cur_corner1, False)
# in the rare case that no transformation were found, use the
# previously known transform
if (T is None):
T = last_T
last_T = T
# decompose the transform
dx = T[0][2]
dy = T[1][2]
da = atan2(T[1][0], T[0][0])
file.write(str(k) + " " + str(dx) + " " + str(dy) + " " + str(da) + "\n")
k += 1
# store the trasnform parameters
prev_to_cur_transform.append(TransformParam(da, dx, dy))
prev = cur
gray_prev = gray_cur
file.close()
print "Done"
return prev_to_cur_transform
tpts1=pts1[ii][0]
tpts2=nextPts[ii][0]
# if status[ii][0]==1:
if (status[ii]==1) and (err[ii]<20):
cv2.circle(frm2c, tuple(tpts1), 3, (255,0,0))
cv2.line(frm2c, tuple(tpts1), tuple(tpts2), (255,255,0))
cv2.circle(frm2c, tuple(tpts2), 3, (0,0,255))
# pts1Good=pts1[ ((status==1) & (err<20))]
pts1Good=pts1[ status==1 ]
pts1Good=np.reshape(pts1Good, (pts1Good.shape[0],1,pts1Good.shape[1]))
# nextPtsG=nextPts[((status==1) & (err<20))]
nextPtsG=nextPts[ status==1 ]
nextPtsG=np.reshape(nextPtsG, (nextPtsG.shape[0],1,nextPtsG.shape[1]))
# print pts1Good.shape
# print nextPtsG.shape
T=cv2.estimateRigidTransform(pts1Good, nextPtsG, False)
dx,dy=T[0,2],T[1,2]
dxy=np.array([dx,dy])
pCenter=np.array([frm2c.shape[1]/2, frm2c.shape[0]/2])
dr=np.sqrt(dx**2+dy**2)
cv2.line(frm2c, (pCenter[0],pCenter[1]), (int(pCenter[0]+dxy[0]),int(pCenter[1]+dxy[1])), (0,255,0))
cv2.circle(frm2c, (pCenter[0],pCenter[1]), 5, (0,255,0))
cv2.circle(frm2c, (pCenter[0],pCenter[1]), int(dr), (0,255,0))
cv2.circle(frm2c, (int(pCenter[0]+dxy[0]),int(pCenter[1]+dxy[1])), 5, (0,0,255))
print T
print "dxy=(%s, %s)" % (T[0,2], T[1,2])
cv2.imshow("win2", frm2c)
while True:
key = cv2.waitKey(0)
if key==27: