def process(self, *mats):
# x = time.perf_counter()
DOWNSIZE_CAMERA = self.options['downsize_camera']
mat = cv2.cvtColor(mats[0], cv2.COLOR_BGR2GRAY)
img2 = resize(
to_umat(mat), int(mat.shape[1] * DOWNSIZE_CAMERA),
int(mat.shape[0] *
DOWNSIZE_CAMERA)) if DOWNSIZE_CAMERA else mat # trainImage
board = self.static_process('upper', 'lower')
# find the keypoints and descriptors with SIFT
kp2, des2 = self.detector.detectAndCompute(img2, None)
cam = {"img": img2, "kp": kp2, "des": des2}
p = resize(mats[0],
int(mats[0].shape[1] * self.options['downsize_camera']),
int(mats[0].shape[0] * self.options['downsize_camera']))
p_mat = p.copy()
p, M = self.match(board, cam, p, (255, 0, 0))
if self.options['show_keypoints']:
p = cv2.drawKeypoints(p, kp2, None, (255, 255, 0))
assert p is not None
self.post_shm(p_mat, p, M)
self.post("outline", p)
def find_key_descriptors(im):
if self.options['source_x_scale_%s' % image] != 0 and self.options[
'source_y_scale_%s' % image] != 0:
scaledim = resize(
im,
int(im.shape[1] *
self.options['source_x_scale_%s' % image]),
int(im.shape[0] *
self.options['source_y_scale_%s' % image]))
scaledim = np.pad(scaledim,
((PADDING, PADDING), (PADDING, PADDING)),
'constant',
constant_values=255)
rx = self.options['source_x_scale_%s' % image]
ry = self.options['source_y_scale_%s' % image]
else:
scaledim = im
rx = 1
ry = 1
kp, des = self.detector.detectAndCompute(scaledim, None)
self.static[image] = {
"name": image,
"org": im,
"img": scaledim,
"rx": rx,
"ry": ry,
"kp": kp,
"des": des
}
self.post(
image,
cv2.drawKeypoints(scaledim, kp, None, (0, 0, 255), flags=0))
return self.static[image]
def process(self, mat):
mat = resize(mat, mat.shape[1] // 4, mat.shape[0] // 4)
col = cv2.cvtColor(mat, cv2.COLOR_BGR2LAB).astype(np.int16)
dst1 = dst_thresh(col,
(self.options['t_l_trg'], self.options['t_a_trg'],
self.options['t_b_trg']))
dst2 = dst_thresh(col,
(self.options['b_l_trg'], self.options['b_a_trg'],
self.options['b_b_trg']))
cdst = dst1.astype(np.uint16) * dst2.astype(np.uint16)
t3 = cdst < (self.options['combined_thresh']**2)
self.post('dst1', dst1)
self.post('dst2', dst2)
self.post('cdst', (cdst**.5).astype(np.uint8))
#t2 = dst < self.options['d_thresh']
#t1 = t1.astype(np.uint8) * 255
#t2 = t2.astype(np.uint8) * 255
t3 = t3.astype(np.uint8) * 255
#self.post('t1', t1)
#self.post('t2', t2)
self.post('t3', t3)
img, contours, hierarchy = cv2.findContours(t3, cv2.RETR_CCOMP,
cv2.CHAIN_APPROX_SIMPLE)
if contours is None or hierarchy is None:
shm.bins_status.lever_visible.set(False)
else:
valid_contours = (hierarchy[0, :, 3] < 0) # (hierarchy[0,:,2] < 0)
#print(valid_contours.shape, contours.shape)
#print(hierarchy)
#print(len(contours), valid_contours)
#contours = contours[valid_contours]
#print([(cv2.contourArea(x), cv2.arcLength(x, True) ** 2 / 2) for x in contours])
contours = [
x for i, x in enumerate(contours)
if valid_contours[i] and cv2.contourArea(x) >
cv2.arcLength(x, True)**2 / 2 * self.options['compactness']
]
#print(contours)
if contours:
mx = max(contours, key=cv2.contourArea)
mm = cv2.moments(mx)
if mm['m00'] != 0:
cx = mm['m10'] / mm['m00']
cy = mm['m01'] / mm['m00']
cv2.drawContours(mat, [mx], -1, (0, 255, 0), 2)
norm_x = cx / mat.shape[1] - .5
norm_y = (cy - mat.shape[0] / 2) / mat.shape[1]
shm.bins_status.lever_x.set(norm_x)
shm.bins_status.lever_y.set(norm_y)
shm.bins_status.lever_sz.set(cv2.contourArea(mx)**.5)
shm.bins_status.lever_visible.set(True)
else:
shm.bins_status.lever_visible.set(False)
else:
shm.bins_status.lever_visible.set(False)
self.post('mat', mat)
def static_process(self, image):
def find_key_descriptors(im):
if self.options['source_x_scale_%s' % image] != 0 and self.options[
'source_y_scale_%s' % image] != 0:
scaledim = resize(
im,
int(im.shape[1] *
self.options['source_x_scale_%s' % image]),
int(im.shape[0] *
self.options['source_y_scale_%s' % image]))
scaledim = np.pad(scaledim,
((PADDING, PADDING), (PADDING, PADDING)),
'constant',
constant_values=255)
rx = self.options['source_x_scale_%s' % image]
ry = self.options['source_y_scale_%s' % image]
else:
scaledim = im
rx = 1
ry = 1
kp, des = self.detector.detectAndCompute(scaledim, None)
self.static[image] = {
"name": image,
"org": im,
"img": scaledim,
"rx": rx,
"ry": ry,
"kp": kp,
"des": des
}
self.post(
image,
cv2.drawKeypoints(scaledim, kp, None, (0, 0, 255), flags=0))
return self.static[image]
if image in self.static:
if self.static[image]['rx'] == self.options['source_x_scale_%s'%image] and \
self.static[image]['ry'] == self.options['source_y_scale_%s'%image]:
return self.static[image]
else:
im = self.static[image]["org"]
return find_key_descriptors(im)
else:
im = simple_gaussian_blur(
cv2.imread('buoy_images/%s.png' % image, 0), 11, 3)
im = resize(im, im.shape[1] // 2, im.shape[0] // 2)
return find_key_descriptors(im)
def crop_by_mask(cvtmat, mask, x, y, r, shrink=False):
import time
t = time.time()
cvtmat = cvtmat[:, :, 1:] # TODO: make this adjustable
print('foo a', time.time() - t)
t = time.time()
mask = mask[y - r:y + r, x - r:x + r]
print('foo b', time.time() - t)
t = time.time()
cropped = cvtmat[y - r:y + r, x - r:x + r]
print('foo c', time.time() - t)
t = time.time()
cropped = cv2.bitwise_and(cropped, cropped, mask=mask)
print('foo d', time.time() - t)
if shrink:
size = min(cropped.shape[0], cropped.shape[1], shrink)
cropped = resize(cropped, size, size)
return cropped
def process(self, mat):
self.post('org', mat)
mat = resize(mat, mat.shape[1]//2, mat.shape[0]//2)
shm.bins_garlic.center_x.set(mat.shape[0]//2)
shm.bins_garlic.center_y.set(mat.shape[1]//2)
cvtmat, split = bgr_to_lab(mat)
self.circles = find_yellow_circle(split,
color=[self.options['yellow_{}'.format(s)] for s in COLORSPACE],
distance=self.options['circle_color_distance'],
erode_kernel=self.options['circle_erode_kernel'],
erode_iterations=self.options['circle_erode_iterations'],
dilate_kernel=self.options['circle_dilate_kernel'],
dilate_iterations=self.options['circle_dilate_iterations'],
min_contour_size=self.options['circle_min_contour_size'],
min_circularity=self.options['circle_min_circularity'],
radius_offset=self.options['garlic_circle_r_offset'])
cv2.drawContours(mat, [c['contour'] for c in self.circles], 0, (255, 0, 0), 10)
for c in self.circles:
cv2.circle(mat, *c['circle'], (0, 255, 0), 10)
self.post('circle', mat)
self.find_red_garlic(cvtmat, split)
def process(self, mat):
t = time.perf_counter()
self.post('org', mat)
mat = resize(mat, mat.shape[1]//2, mat.shape[0]//2)
shm.recovery_vampire.cam_x.set(mat.shape[1]//2)
shm.recovery_vampire.cam_y.set(mat.shape[0]//2)
# tt = time.perf_counter()
# print('1 %f' % (tt - t))
# print(mat.shape)
_, split = bgr_to_lab(mat)
d = self.options['vampire_color_distance']
color = [self.options["yellow_%s" % c] for c in COLORSPACE]
self.rectangles = self.find_yellow_rectangle(split, color, d, self.options['erode_kernel_size'],
self.options['erode_iterations'],
self.options['dilate_kernel_size'],
self.options['dilate_iterations'],
self.options['contour_size_min'],
self.options['rectangularity_thresh'],
-self.options['rectangle_padding'])
# t = time.perf_counter()
# print('2 %f' % (t - tt))
for y in self.rectangles:
rectangle = cv2.boxPoints(y['rectangle'])
mat = cv2.drawContours(mat, [np.int0(rectangle)], 0, (0, 0, 255), 10)
color = [self.options["purple_%s" % c] for c in COLORSPACE]
# purple = self.find_color(mat, color, d, use_first_channel=False, erode_mask=True, dilate_mask=True, iterations=3, rectangular=False)
# self.post('purple', purple)
# purple_contours = self.contours_and_filter(purple, self.options['contour_size_min'])
self.find_vampire(mat, split, color, d)
# tt = time.perf_counter()
# print('3 %f' % (tt-t))
mat = cv2.drawContours(mat, [r['contour'] for r in self.rectangles], -1, (0, 255, 0), 10)
# mat = cv2.drawContours(mat, purple_contours, -1, (0, 255, 0), 10)
self.post('yellow_contours', mat)
def find_key_descriptors(im):
if self.options['source_x_scale_%s' % image] != 0 and self.options[
'source_y_scale_%s' % image] != 0:
scaledim = resize(
im,
int(im.shape[1] *
self.options['source_x_scale_%s' % image]),
int(im.shape[0] *
self.options['source_y_scale_%s' % image]))
scaledim = np.pad(scaledim,
((PADDING, PADDING), (PADDING, PADDING)),
'constant',
constant_values=255)
rx = self.options['source_x_scale_%s' % image]
ry = self.options['source_y_scale_%s' % image]
else:
scaledim = im
rx = 1
ry = 1
scaledim = simple_gaussian_blur(scaledim, BLUR_KERNEL, BLUR_SD)
kp, des = self.detector.detectAndCompute(scaledim, None)
self.static[image] = {
"name": image,
"org": im,
"img": scaledim,
"rx": rx,
"ry": ry,
"kp": kp,
"des": des,
"separation": self.options['board_separation']
}
keypoints = cv2.drawKeypoints(scaledim.copy(),
kp,
None, (0, 0, 255),
flags=0)
self.post(image, keypoints)
return self.static[image]
def find_key_descriptors(im):
scaledim = resize(
im,
int(im.shape[1] * self.options['source_x_scale_%s' % image]),
int(im.shape[0] * self.options['source_y_scale_%s' % image]))
#mean = 0
#sigma = 20
#gauss = np.random.normal(mean,sigma,scaledim.shape)
#scaledim = scaledim.astype(np.int16) + gauss.astype(np.int8)
#np.clip(scaledim, 0, 255, out=scaledim)
#scaledim = scaledim.astype(np.uint8)
scaledim = np.pad(scaledim,
((PADDING, PADDING), (PADDING, PADDING)),
'constant',
constant_values=255)
rx = self.options['source_x_scale_%s' % image]
ry = self.options['source_y_scale_%s' % image]
scaledim = simple_gaussian_blur(scaledim, BLUR_KERNEL, BLUR_SD)
kp, des = self.detector.detectAndCompute(scaledim, None)
self.static[image] = {
"name": image,
"org": im,
"img": scaledim,
"rx": rx,
"ry": ry,
"kp": kp,
"des": des
}
keypoints = cv2.drawKeypoints(scaledim.copy(),
kp,
None, (0, 0, 255),
flags=0)
self.post(image, keypoints)
return self.static[image]
def process(self, *mats):
x = time.perf_counter()
DOWNSIZE_CAMERA = self.options['downsize_camera']
img2 = resize(
to_umat(mats[0]), int(mats[0].shape[1] * DOWNSIZE_CAMERA),
int(mats[0].shape[0] *
DOWNSIZE_CAMERA)) if DOWNSIZE_CAMERA else mats[0] # trainImage
#img2 = cv2.copyMakeBorder(img2, PADDING,PADDING,PADDING,PADDING, cv2.BORDER_REPLICATE)
draugr = self.static_process('draugr')
vetalas = self.static_process('vetalas')
aswang = self.static_process('aswang')
jiangshi = self.static_process('jiangshi')
# find the keypoints and descriptors with SIFT
kp2, des2 = self.detector.detectAndCompute(img2, None)
cam = {"img": img2, "kp": kp2, "des": des2}
p = self.match(draugr, cam, None, (255, 0, 0))
p = self.match(vetalas, cam, p, (0, 255, 0))
p = self.match(aswang, cam, p, (0, 0, 255))
p = self.match(jiangshi, cam, p, (255, 0, 255))
if self.options['show_keypoints']:
p = cv2.drawKeypoints(p, kp2, None, (255, 255, 0))
p = from_umat(p)
self.post_shm()
shm.vamp_buoy_results.camera_x.set(p.shape[1] // 2)
shm.vamp_buoy_results.camera_y.set(p.shape[0] // 2)
self.post("outline", p)
print(time.perf_counter() - x)
def process(self, *mats):
results = shm.gate_vision.get()
h, w, _ = mats[0].shape
h = int(h * self.options['resize_height_scale'])
w = int(w * self.options['resize_width_scale'])
results.img_height = h
results.img_width = w
mat = resize(mats[0], w, h)
#print(np.mean(mat))
avg_brightness_ratio = np.mean(mat) / REFERENCE_BRIGHTNESS
nonblack_thresh_dist = self.options['nonblack_thresh'] * avg_brightness_ratio
lab, lab_split = bgr_to_lab(mat)
median_a = np.median(lab_split[1])
median_b = np.median(lab_split[2])
median_filter_a = range_threshold(lab_split[1], median_a - self.options['water_a_thresh'], median_a + self.options['water_a_thresh'])
median_filter_b = range_threshold(lab_split[2], median_b - self.options['water_b_thresh'], median_b + self.options['water_b_thresh'])
if self.options['debug']:
self.post('median filter a', median_filter_a)
self.post('median filter b', median_filter_b)
nonwater_mask, _ = gray_to_bgr(255 - (median_filter_a & median_filter_b))
self.post('nonwater', nonwater_mask)
# Tuned for a 320x256 image
vehicle_depth = shm.kalman.depth.get()
reflection_cutoff = min(h, int(max(0, 3 - vehicle_depth)**2 * CUTOFF_SCALAR))
mat[:reflection_cutoff] *= 0
tmp = mat.copy()
draw_text(tmp, 'Depth: {:.2f}'.format(vehicle_depth), (30, 30), 0.5, color=(255, 255, 255))
self.post('mat', tmp)
#lab, lab_split = bgr_to_lab(mat)
#nonblack_mask, _ = gray_to_bgr(np.uint8(255 * (lab_split[0] > self.options['nonblack_thresh'])))
nonblack_mask, _ = gray_to_bgr(np.uint8(255 * (np.var(mat, axis=2) > nonblack_thresh_dist)))
self.post('nonblack', nonblack_mask)
mat &= nonblack_mask
mat &= nonwater_mask
mat = to_umat(mat)
mat = simple_gaussian_blur(mat, to_odd(self.options['blur_kernel']),
self.options['blur_std'])
lab, lab_split = bgr_to_lab(mat)
threshed, dists = thresh_color_distance([lab_split[0], lab_split[1], lab_split[2]],
[self.options['lab_l_ref'], self.options['lab_a_ref'],
self.options['lab_b_ref']],
self.options['color_dist_thresh'], auto_distance_percentile=self.options['auto_distance_percentile'],
ignore_channels=[0], weights=[2, 0, 15])
if self.options['debug']:
self.post('threshed', threshed)
self.post('dists', dists)
dilated = dilate(threshed, rect_kernel(self.options['dilate_kernel']))
if self.options['debug']:
self.post('dilated', dilated)
eroded = erode(dilated, rect_kernel(self.options['erode_kernel']))
if self.options['debug']:
self.post('eroded', eroded)
contours = outer_contours(eroded)
areas = [*map(contour_area, contours)]
centroids = [*map(contour_centroid, contours)]
xs = [c[0] for c in centroids]
ys = [c[1] for c in centroids]
rects = [*map(min_enclosing_rect, contours)]
lengths = [max(r[1]) for r in rects]
ratios = [max(r[1]) / (1e-30 + min(r[1])) for r in rects]
vehicle_roll = shm.kalman.roll.get()
lines = [cv2.fitLine(c, cv2.DIST_L2, 0, 0.01, 0.01) for c in contours]
angles = [np.degrees(np.arctan2(line[1], line[0]))[0] for line in lines]
angles = [min(abs(90 - a - vehicle_roll), abs(-90 - a - vehicle_roll)) for a in angles]
rectangularities = [a / (1e-30 + rect[1][0] * rect[1][1]) for (c, a, rect) in zip(contours, areas, rects)]
contours = [ContourFeats(*feats) for feats in zip(contours, areas, xs, ys, rectangularities, angles, lengths, ratios)]
contours = [*filter(lambda c: c.area > self.options['min_contour_area'], contours)]
self.post_contours('area', h, w, contours)
contours = [*filter(lambda c: c.angle < self.options['max_angle_from_vertical'], contours)]
self.post_contours('angle', h, w, contours)
contours = [*filter(lambda c: c.length > self.options['min_length'], contours)]
self.post_contours('length', h, w, contours)
#contours = [*filter(lambda c: c.rect > self.options['min_contour_rect'], contours)]
#self.post_contours('rect', h, w, contours)
contours = [*filter(lambda c: c.ratio > self.options['min_contour_ratio'], contours)]
self.post_contours('ratio', h, w, contours)
contours = sorted(contours, key=lambda c: c.area)[:6]
contours_by_x = sorted(contours, key=lambda c: c.x)
contours_by_x = filter_duplicates_sorted_by_x(contours_by_x)
leftmost = try_index(contours_by_x, 0)
middle = try_index(contours_by_x, 1)
rightmost = try_index(contours_by_x, 2)
tmp = np.zeros((h, w, 3))
results.leftmost_visible = leftmost is not None
results.middle_visible = middle is not None
results.rightmost_visible = rightmost is not None
draw_text(tmp, 'Roll: {:.2f}'.format(vehicle_roll), (30, 30), 0.5, color=(255, 255, 255))
if leftmost is not None:
draw_contours(tmp, [leftmost.contour], color=(255, 0, 0), thickness=-1)
draw_circle(tmp, (leftmost.x, leftmost.y), 5, color=(255, 255, 255), thickness=-1)
results.leftmost_x = leftmost.x
results.leftmost_y = leftmost.y
results.leftmost_len = leftmost.length
if middle is not None:
draw_contours(tmp, [middle.contour], color=(0, 255, 0), thickness=-1)
draw_circle(tmp, (middle.x, middle.y), 5, color=(255, 255, 255), thickness=-1)
results.middle_x = middle.x
results.middle_y = middle.y
results.middle_len = middle.length
if rightmost is not None:
draw_contours(tmp, [rightmost.contour], color=(0, 0, 255), thickness=-1)
draw_circle(tmp, (rightmost.x, rightmost.y), 5, color=(255, 255, 255), thickness=-1)
results.rightmost_x = rightmost.x
results.rightmost_y = rightmost.y
results.rightmost_len = rightmost.length
shm.gate_vision.set(results)
self.post('contours', tmp)
def process(self, mat):
camera_scale = self.options['camera_scale']
debug = self.options['debug']
if camera_scale != 0:
mat = resize(mat, int(mat.shape[1] * camera_scale),
int(mat.shape[0] * camera_scale))
l_mat = cv2.cvtColor(mat, cv2.COLOR_BGR2Lab)
if debug:
mm = l_mat.astype(np.int32)
dst_lid = (mm[:,:,0] - self.options['lid_l_trg']) ** 2 + \
(mm[:,:,1] - self.options['lid_a_trg']) ** 2 + \
(mm[:,:,2] - self.options['lid_b_trg']) ** 2
self.post('yellowness_lid', ((dst_lid / 3)**.5).astype(np.uint8))
#np.clip(dst_lid, 0, 255, out=dst_lid)
#dst_lid = dst_lid.astype(np.uint8)
#res, yellow_mask_lid = cv2.threshold(to_umat(dst_lid), self.options['lid_d_thresh'], 255, cv2.THRESH_BINARY_INV)
yellow_mask_lid = (dst_lid < self.options['lid_d_thresh']**
2).astype(np.uint8) * 255
self.post('yellow_mask_lid', yellow_mask_lid)
img2 = cv2.cvtColor(to_umat(mat), cv2.COLOR_BGR2GRAY)
edg = cv2.Canny(img2,
self.options['canny1'],
self.options['canny2'],
apertureSize=3)
#yellow_edg_msk = cv2.erode(yellow_mask_lid, kernel)
#yellow_edg_msk = cv2.dilate(yellow_edg_msk, kernel, iterations=4)
#self.post('edg0', edg)
#edg = cv2.bitwise_and(edg, yellow_edg_msk)
self.post('edg', edg)
lines = cv2.HoughLines(edg, 1, np.pi / 180,
self.options['houghness']).get()
clrs = [(0, 0, 255), (0, 255, 0), (255, 0, 0), (0, 255, 255),
(255, 255, 0), (255, 0, 255), (0, 0, 128), (0, 128, 0),
(128, 0, 0)]
if lines is None:
self.post('lines', mat)
shm.bins_status.cover_visible.set(False)
return
lines[lines[:, 0, 0] < 0, :, 1] += np.pi
lines[:, :, 0] = np.abs(lines[:, :, 0])
lvecs = np.exp(1j * lines[:, :, 1])
mlvecs = lvecs**2
#print(lvecs)
if debug:
for i, (dst, vect) in enumerate(zip(lines[:, 0, 0], lvecs[:, 0])):
ctr = vect * dst
vc = vect * 1j
p1 = ctr + 1000 * vc
p2 = ctr - 1000 * vc
mat = cv2.line(mat, (int(p1.real), int(p1.imag)),
(int(p2.real), int(p2.imag)), (0, 0, 200), 2)
#print(vect, dst)
m = cv2.BFMatcher(cv2.NORM_L2)
cc = mlvecs.astype(np.complex64).view(np.float32)
centers = []
if len(mlvecs) > 1:
dsst = 2.83 - shm.kalman.depth.get()
plen = 75 / dsst if dsst > 0 else 1000
res = m.match(cc, -cc)
for m in res:
if m.distance > .05: continue
d1 = lvecs[m.trainIdx, 0], lines[m.trainIdx, 0, 0]
d2 = lvecs[m.queryIdx, 0], lines[m.queryIdx, 0, 0]
try:
center = np.linalg.solve(
np.complex64([d1[0] / abs(d1[0]), d2[0] / abs(d2[0])
]).view(np.float32).reshape(2, -1),
[[d1[1]], [d2[1]]])[:, 0].astype(
np.float64) #.view(np.complex64)[0]
except np.linalg.linalg.LinAlgError:
print('singular matrix')
continue
cC = complex(*center)
mz = np.zeros(mat.shape[:-1], dtype=np.uint8)
mz2 = np.zeros(mat.shape[:-1], dtype=np.uint8)
poly = make_poly(d1[0], d2[0])
rpoly = make_poly(d1[0], -d2[0])
mz = cv2.fillConvexPoly(mz, (cC + plen * poly).view(
np.float32).astype(np.int32), 255)
mz = cv2.fillConvexPoly(mz, (cC - plen * poly).view(
np.float32).astype(np.int32), 255)
mz2 = cv2.fillConvexPoly(mz2, (cC + plen * rpoly).view(
np.float32).astype(np.int32), 255)
mz2 = cv2.fillConvexPoly(mz2, (cC - plen * rpoly).view(
np.float32).astype(np.int32), 255)
m1, sd1 = cv2.meanStdDev(img2, mask=mz)
m2, sd2 = cv2.meanStdDev(img2, mask=mz2)
m1, sd1, m2, sd2 = (x.get()[0, 0] for x in (m1, sd1, m2, sd2))
score = abs(m1 - m2) / (sd1 * sd2)
if score < self.options['min_cross_score']: continue
no_swap = np.cross(
*np.complex128([[d1[0]], [d2[0]]]).view(np.float64)) > 0
d1, d2 = (d1, d2) if no_swap else (d2, d1)
long_axis, short_axis = (d1[0], d2[0]) if m1 < m2 else (d2[0],
d1[0])
sm = mz if (m1 > m2) else mz2 # ^ no_swap
m_color = cv2.mean(l_mat, mask=sm)[0:3]
passes_color = cv2.norm(
np.float32(m_color),
np.float32([
self.options['lid_l_trg'], self.options['lid_a_trg'],
self.options['lid_b_trg']
])) < self.options['lid_d_thresh']
if passes_color:
centers.append(
(cC, d1, d2, score, long_axis, short_axis, sm))
try:
mat = cv2.circle(mat, (int(center[0]), int(center[1])), 10,
(255, 0, 0) if passes_color else
(0, 200, 200), 2)
except (ValueError, OverflowError):
pass
if centers:
mx = max(centers, key=lambda h: h[3])
center, d1, d2, score, long_axis, short_axis, sm = mx
#dsst = 4.26 - shm.kalman.depth.get() - .762
pts = (np.complex64([
long_axis + short_axis, long_axis - short_axis,
-long_axis - short_axis, -long_axis + short_axis
])[:, np.newaxis] * plen + center).view(np.float32).astype(
np.int32)
print(pts)
mat = cv2.polylines(mat, [pts], True, (255, 255, 255))
print(score, shm.kalman.depth.get())
if long_axis.imag < 0: long_axis *= -1
if short_axis.real < 0: short_axis *= -1
#self.post('sm', sm)
shm.bins_status.cover_x.set(center.real / mat.shape[1] - .5)
shm.bins_status.cover_y.set(
(center.imag - mat.shape[0] / 2) / mat.shape[1])
shm.bins_status.cover_maj_x.set(long_axis.real)
shm.bins_status.cover_maj_y.set(long_axis.imag)
shm.bins_status.cover_min_x.set(short_axis.real)
shm.bins_status.cover_min_y.set(short_axis.imag)
shm.bins_status.cover_visible.set(True)
for ax, clr in ((short_axis, (255, 0, 0)), (long_axis, (0, 255,
0))):
p1 = center + ax * 1000
p2 = center - ax * 1000
mat = cv2.line(mat, (int(p1.real), int(p1.imag)),
(int(p2.real), int(p2.imag)), clr, 2)
mat = cv2.circle(mat, (int(center.real), int(center.imag)), 10,
(0, 255, 0), 2)
else:
shm.bins_status.cover_visible.set(False)
self.post('lines', mat)
def process(self, *mats):
# x = time.perf_counter()
CAMERA_SCALE = self.options['camera_scale']
t = time.perf_counter()
print('a',
time.perf_counter() - t)
t = time.perf_counter()
mat = resize(mats[0], int(mats[0].shape[1] * CAMERA_SCALE),
int(mats[0].shape[0] *
CAMERA_SCALE)) if CAMERA_SCALE else mats[0]
temp = mat.astype(
np.uint16) * 2 #self.options_dict['PPX_contrast'].value
mat = np.clip(temp, 0, 255).astype(np.uint8)
p = mat.copy()
#rv, ccs = cv2.findChessboardCorners(mat, (1, 1))
#print(rv)
#l_mat = cv2.cvtColor(mat, cv2.COLOR_BGR2Lab)
#mm = l_mat.astype(np.int16)
#dst = np.abs(mm[:,:,0] - self.options['img_l_trg']) + \
# np.abs(mm[:,:,1] - self.options['img_a_trg']) + \
# np.abs(mm[:,:,2] - self.options['img_b_trg'])
#self.post('yellowness', (dst // 3).astype(np.uint8))
#np.clip(dst, 0, 255, out=dst)
#dst = dst.astype(np.uint8)
#res, yellow_mask = cv2.threshold(dst, self.options['img_d_thresh'], 255, cv2.THRESH_BINARY_INV)
#self.post('yellow_mask', yellow_mask)
img2 = cv2.cvtColor(to_umat(mat), cv2.COLOR_BGR2GRAY)
print('b',
time.perf_counter() - t)
t = time.perf_counter()
#corners = cv2.cornerHarris(img2, 2, 3, .04)
#print(corners.get().dtype)
#cg = corners.get()
#cg[cg <= 0] = cg[cg > 0].min()
#print(cg.max())
#print(cg.min(), cg.max())
#ll = np.log(cg)
#print(ll.min(), ll.max())
#self.post('harris', np.clip((ll * 5 + 128), 0, 255).astype(np.uint8))
#img2 = to_umat(np.pad(cv2.cvtColor(mat, cv2.COLOR_BGR2GRAY), ((PADDING, PADDING), (PADDING, PADDING)), 'constant', constant_values=255))
#mat = to_umat(np.pad(img2.get(), ((PADDING, PADDING), (PADDING, PADDING)), 'constant', constant_values=255))
#img2 = resize(to_umat(mat), int(mat.shape[1]*camera_scale), int(mat.shape[0]*camera_scale)) if DOWNSIZE_CAMERA else mat # trainImage
print('h',
time.perf_counter() - t)
t = time.perf_counter()
bat = self.static_process('bat')
wolf = self.static_process('wolf')
print('i',
time.perf_counter() - t)
t = time.perf_counter()
#black_areas = img2 < self.options['min_gray']
res, black_areas = cv2.threshold(img2, self.options['min_gray'], 255,
cv2.THRESH_BINARY_INV)
black_areas = cv2.erode(black_areas, kernel)
black_areas = cv2.dilate(black_areas, kernel, iterations=2)
self.post('black_areas', black_areas)
img, contours, hierarchy = cv2.findContours(black_areas,
cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE)
msk = np.zeros(mat.shape[:-1], dtype=np.uint8)
#szcs = sorted(contours, key=cv2.contourArea, reverse=True)
print('j',
time.perf_counter() - t)
t = time.perf_counter()
for i, x in enumerate(contours): #szcs[:2]:
pts = cv2.boxPoints(cv2.minAreaRect(x))
#mm = np.mean(x, axis=0)
mpp = np.mean(pts, axis=0)
#print(pts)
#print(np.mean(pts, axis=0))
cv2.drawContours(msk,
[np.int0(pts - (pts - (pts * 4 + mpp) / 5) * .1)],
-1, i + 1, -1)
#print(x * 9 + mm, mm)
# print(x)
#cv2.drawContours(msk, [np.int0((x * 2 + mm) / 3)], -1, i+1, -1)
#for p in szcs[0:2]:
# msk = cv2.fillPoly(msk, p, 255)
colors = np.uint8([(0, 0, 0), (0, 0, 255), (0, 255, 0), (255, 0, 0),
(0, 255, 255), (255, 255, 0), (255, 0, 255),
(255, 255, 255)])
self.post('blk_fill', colors[np.clip(msk, 0, 7)])
print('k',
time.perf_counter() - t)
t = time.perf_counter()
#print(szcs)
#black_areas = black_areas.astype(np.uint8)
#target_area = cv2.dilate(black_areas, kernel, iterations=10)
#target_area = target_area.get()
#self.post('target_area', target_area)
#target_area = cv2.erode(yellow_mask, kernel)
#target_area = cv2.dilate(target_area, kernel, iterations=10)
#target_area &= msk
target_area = msk
#self.post('target_area', target_area)
# find the keypoints and descriptors with SIFT
#img2 = cv2.UMat(np.pad(img2.get(), ((PADDING, PADDING), (PADDING, PADDING)), 'constant', constant_values=255)) # boo
print('l',
time.perf_counter() - t)
t = time.perf_counter()
kp2, des2 = self.detector.detectAndCompute(img2, None)
print('l1',
time.perf_counter() - t)
t = time.perf_counter()
if des2 is None:
print('No points found')
return
dg = des2.get()
if dg is None:
print('No points found')
return
#cv2.UMat(des2, [0, 1, 2])
#print(dir(des2))
#print(des2.get([0, 1, 2]))
#print(kp2)
#print(max(x.pt[1] for x in kp2))
#print(max(x.pt[0] for x in kp2))
#p = resize(mats[0], int(mats[0].shape[1] * self.options['camera_scale']), int(mats[0].shape[0] * self.options['camera_scale']))
#p = cv2.copyMakeBorder(p, PADDING, PADDING, PADDING, PADDING, cv2.BORDER_CONSTANT, None, (255, 255, 255))
if self.options['show_keypoints']:
p = cv2.drawKeypoints(p, kp2, None, (0, 255, 255))
#print([x.pt for x in kp2])
idxs = [
i for (i, x) in enumerate(kp2)
if (0 < x.pt[1] < target_area.shape[0]) and
(0 < x.pt[0] < target_area.shape[1]) and target_area[int(x.pt[1]),
int(x.pt[0])]
]
kp2 = [kp2[i] for i in idxs]
des2 = cv2.UMat(dg[idxs])
#print(kp2[0].pt)
cam = {"img": img2, "kp": kp2, "des": des2}
if self.options['show_keypoints']:
p = cv2.drawKeypoints(p, kp2, None, (255, 255, 0))
p_mat = p.copy()
def get_center_ang(img, mat):
ii = img['img']
pts = np.float32([[ii.shape[1] / 2, ii.shape[0] / 2],
[ii.shape[1] / 2, 0]]).reshape(-1, 1,
2) + PADDING
middle, top = cv2.perspectiveTransform(pts, mat)[:, 0, :]
up_vec = top - middle
return middle, np.arctan2(up_vec[1], up_vec[0])
if kp2:
print('m',
time.perf_counter() - t)
t = time.perf_counter()
p, M1 = self.match(bat, cam, p, (0, 0, 255), msk, len(contours))
p, M2 = self.match(wolf, cam, p, (0, 255, 0), msk, len(contours))
if M1 is not None:
ctr, ang = get_center_ang(bat, M1)
shm.bins_status.bat_x.set(ctr[0] / mat.shape[1] - .5)
shm.bins_status.bat_y.set(
(ctr[1] - mat.shape[0] / 2) / mat.shape[1])
shm.bins_status.bat_angle.set(ang)
shm.bins_status.bat_visible_frames.set(
shm.bins_status.bat_visible_frames.get() + 1)
shm.bins_status.bat_visible.set(M1 is not None)
if M2 is not None:
ctr, ang = get_center_ang(bat, M2)
shm.bins_status.wolf_x.set(ctr[0] / mat.shape[1] - .5)
shm.bins_status.wolf_y.set(
(ctr[1] - mat.shape[0] / 2) / mat.shape[1])
shm.bins_status.wolf_angle.set(ang)
shm.bins_status.wolf_visible_frames.set(
shm.bins_status.wolf_visible_frames.get() + 1)
shm.bins_status.wolf_visible.set(M2 is not None)
print('n',
time.perf_counter() - t)
t = time.perf_counter()
assert p is not None
#try:
# self.post_shm(p_mat, p, M)
#except cv2.error as e:
# print(e)
self.post("outline", p)
def match(self, img, min_match=10, ratio=0.7, draw=False):
"""
Find all instances of source images in an image.
img: The image to be matched.
min_match: The minimum number of keypoint matches to identify a
source. Increasing this number allows sources to be
identified more accurately, but less reliably.
ratio: A number [0..1] that is used to identify a "good" match
using the ratio test. A higher number means more matches
are considered "good", but is also more likely to be noise.
draw: Whether to draw the matches. I really believe it is not
required, usually it is possible to just compare the
keypoints through human eye, but is a useful tool for
debugging.
The function returns the following as a tuple in this order:
- A list of identified sources, which is a tuple including the
following in this order:
- the name of the source
- the good matches found
- a single contour specifying the area of the match as a
transformed rectangle,
- a mask of the matched area,
- An image showing all matches to the source image;
- The keypoints of the imaged passed
- The feature descriptors of the image passed
"""
kp, des = self.sift.detectAndCompute(img, None)
matched = []
draw_params = dict(matchColor=(0, 255, 0),
singlePointColor=None,
matchesMask=None,
flags=2)
for name, val in self.sources.items():
matches = self.matcher.knnMatch(val["des"], des, k=2)
good = SIFT._ratio_test(matches, ratio=ratio)
if len(good) < min_match:
continue
# TODO: What if we don't do homogenous transform?
src_pts = np.float32([val["kp"][m.queryIdx].pt for m in good])\
.reshape(-1,1,2)
dst_pts = np.float32([kp[m.trainIdx].pt for m in good])\
.reshape(-1,1,2)
matrix, mask = cv2.findHomography(src_pts, dst_pts, cv2.RANSAC,
5.0)
matchesMask = mask.ravel().tolist()
h, w = val["source"].shape
pts = np.float32([[0, 0], [0, h - 1], [w - 1, h - 1],
[w - 1, 0]]).reshape(-1, 1, 2)
try:
# OpenCV sometimes throws assertion errors here. It should have
# been fixed by updating opencv, but it's here just in case
dst = np.int32(cv2.perspectiveTransform(pts, matrix))
except cv2.error as e:
print(e)
continue
drawim = None
if draw:
draw_params["matchesMask"] = matchesMask
drawim = cv2.drawMatches(val["source"], val["kp"], img, kp,
good, None, **draw_params)
drawim = resize(drawim, int(drawim.shape[1] * 0.5),
int(drawim.shape[0] * 0.5))
matched.append((val["name"], good, dst, matchesMask, drawim))
return matched, kp, des