def show3DCloud(ff, stripe, detected_line=None, stripe_points=None, config=None):
""" Shows 3D point cloud of stripe points.
"""
if detected_line is None:
detected_line = []
stripe_indices = stripe.nonzero()
xs = stripe_indices[1]
ys = stripe_indices[0]
zs = ff.maxframe[stripe_indices]
logDebug('points: ', len(xs))
# Plot points in 3D
fig = plt.figure()
ax = fig.add_subplot(111, projection='3d')
ax.scatter(xs, ys, zs)
if detected_line and len(stripe_points):
xs = [detected_line[0][1], detected_line[1][1]]
ys = [detected_line[0][0], detected_line[1][0]]
zs = [detected_line[0][2], detected_line[1][2]]
ax.plot(ys, xs, zs, c = 'r')
x1, x2 = ys
y1, y2 = xs
z1, z2 = zs
detected_points = getAllPoints(stripe_points, x1, y1, z1, x2, y2, z2, config,
fireball_detection=False)
if detected_points.any():
detected_points = np.array(detected_points)
xs = detected_points[:,0]
ys = detected_points[:,1]
zs = detected_points[:,2]
ax.scatter(xs, ys, zs, c = 'r', s = 40)
# Set limits
plt.xlim((0, ff.ncols))
plt.ylim((0, ff.nrows))
ax.set_zlim((0, 255))
plt.show()
plt.clf()
plt.close()
# detected_line = detected_line[0]
xs = [detected_line[0][0], detected_line[1][0]]
ys = [detected_line[0][1], detected_line[1][1]]
zs = [detected_line[0][2], detected_line[1][2]]
ax.plot(ys, xs, zs, c=ln_colors[i % 6])
# Plot grouped points
for i, detected_line in enumerate(line_list):
x1, x2 = detected_line[0][0], detected_line[1][0]
y1, y2 = detected_line[0][1], detected_line[1][1]
z1, z2 = detected_line[0][2], detected_line[1][2]
detected_points = getAllPoints(event_points, x1, y1,
z1, x2, y2, z2, config)
if not detected_points.any():
continue
detected_points = np.array(detected_points)
xs = detected_points[:, 1]
ys = detected_points[:, 0]
zs = detected_points[:, 2]
ax.scatter(xs, ys, zs, c=ln_colors[i % 6], s=40)
print('Elapsed time: ', elapsed_time)
plt.show()
def detectMeteors(ff_directory, ff_name, config, flat_struct=None):
""" Detect meteors on the given FF bin image. Here are the steps in the detection:
- input image (FF bin format file) is thresholded (converted to black and white)
- several morphological operations are applied to clean the image
- image is then broken into several image "windows" (these "windows" are reconstructed from the input FF file, given
an input frame range (e.g. 64-128) which helps reduce the noise further)
- on each "window" the Kernel-based Hough transform is performed to find any lines on the image
- similar lines are joined
- stripe around the lines is extracted
- 3D line finding (third dimension is time) is applied to check if the line propagates in time
- centroiding is performed, which calculates the position and intensity of meteor on each frame
Arguments:
ff_directory: [string] an absolute path to the input FF bin file
ff_name: [string] file name of the FF bin file on which to run the detection on
config: [config object] configuration object (loaded from the .config file)
Keyword arguments:
flat_struct: [Flat struct] Structure containing the flat field. None by default.
Return:
meteor_detections: [list] a list of detected meteors, with these elements:
- rho: [float] meteor line distance from image center (polar coordinates, in pixels)
- theta: [float] meteor line angle from image center (polar coordinates, in degrees)
- centroids: [list] [frame, X, Y, level] list of meteor points
"""
t1 = time()
t_all = time()
# Load the FF bin file
ff = FFfile.read(ff_directory, ff_name)
# Load the mask file
mask = MaskImage.loadMask(config.mask_file)
# Mask the FF file
ff = MaskImage.applyMask(ff, mask, ff_flag=True)
# Apply the flat to maxpixel and avepixel
if flat_struct is not None:
ff.maxpixel = Image.applyFlat(ff.maxpixel, flat_struct)
ff.avepixel = Image.applyFlat(ff.avepixel, flat_struct)
# At the end, a check that the detection has a surface brightness above the background will be performed.
# The assumption here is that the peak of the meteor should have the intensity which is at least
# that of a patch of 4x4 pixels that are of the mean background brightness
min_patch_intensity = 4 * 4 * (np.mean(ff.maxpixel - ff.avepixel) +
config.k1_det * np.mean(ff.stdpixel) +
config.j1)
# # Show the maxpixel image
# show2(ff_name+' maxpixel', ff.maxpixel)
# Get lines on the image
line_list = getLines(ff, config.k1_det, config.j1_det, config.time_slide,
config.time_window_size, config.max_lines_det,
config.max_white_ratio, config.kht_lib_path)
logDebug('List of lines:', line_list)
# Init meteor list
meteor_detections = []
# Only if there are some lines in the image
if len(line_list):
# Join similar lines
line_list = mergeLines(line_list, config.line_min_dist, ff.ncols,
ff.nrows)
logDebug('Time for finding lines:', time() - t1)
logDebug('Number of KHT lines: ', len(line_list))
logDebug(line_list)
# Plot lines
# plotLines(ff, line_list)
# Threshold the image
img_thres = thresholdImg(ff, config.k1_det, config.j1_det)
filtered_lines = []
# Analyze stripes of each line
for line in line_list:
rho, theta, frame_min, frame_max = line
logDebug('rho, theta, frame_min, frame_max')
logDebug(rho, theta, frame_min, frame_max)
# Bounded the thresholded image by min and max frames
img = selectFrames(np.copy(img_thres), ff, frame_min, frame_max)
# Remove lonely pixels
img = morph.clean(img)
# Get indices of stripe pixels around the line
stripe_indices = getStripeIndices(rho, theta, config.stripe_width,
img.shape[0], img.shape[1])
# Extract the stripe from the thresholded image
stripe = np.zeros((ff.nrows, ff.ncols), np.uint8)
stripe[stripe_indices] = img[stripe_indices]
# Show stripe
#COMMENTED
# show2("stripe", stripe*255)
# Show 3D could
# show3DCloud(ff, stripe)
# Get stripe positions
stripe_positions = stripe.nonzero()
xs = stripe_positions[1]
ys = stripe_positions[0]
zs = ff.maxframe[stripe_positions]
# Limit the number of points to search if too large
if len(zs) > config.max_points_det:
# Extract weights of each point
maxpix_elements = ff.maxpixel[ys, xs].astype(np.float64)
weights = maxpix_elements / np.sum(maxpix_elements)
# Random sample the point, sampling is weighted by pixel intensity
indices = np.random.choice(len(zs),
config.max_points_det,
replace=False,
p=weights)
ys = ys[indices]
xs = xs[indices]
zs = zs[indices]
# Make an array to feed into the gropuing algorithm
stripe_points = np.vstack((xs, ys, zs))
stripe_points = np.swapaxes(stripe_points, 0, 1)
# Sort stripe points by frame
stripe_points = stripe_points[stripe_points[:, 2].argsort()]
t1 = time()
logDebug('finding lines...')
# Find a single line in the point cloud
detected_line = find3DLines(stripe_points,
time(),
config,
fireball_detection=False)
logDebug('time for GROUPING: ', time() - t1)
# Extract the first and only line if any
if detected_line:
detected_line = detected_line[0]
# logDebug(detected_line)
# Show 3D cloud
# show3DCloud(ff, stripe, detected_line, stripe_points, config)
# Add the line to the results list
filtered_lines.append(detected_line)
# Merge similar lines in 3D
filtered_lines = merge3DLines(filtered_lines, config.vect_angle_thresh)
logDebug('after filtering:')
logDebug(filtered_lines)
for detected_line in filtered_lines:
# Get frame range
frame_min = detected_line[4]
frame_max = detected_line[5]
# Check if the line covers a minimum frame range
if (abs(frame_max - frame_min) + 1 <
config.line_minimum_frame_range_det):
continue
# Extand the frame range for several frames, just to be sure to catch all parts of a meteor
frame_min -= config.frame_extension
frame_max += config.frame_extension
# Cap values to 0-255
frame_min = max(frame_min, 0)
frame_max = min(frame_max, 255)
logDebug(detected_line)
# Get coordinates of 2 points that describe the line
x1, y1, z1 = detected_line[0]
x2, y2, z2 = detected_line[1]
# Convert Cartesian line coordinates to polar
rho, theta = getPolarLine(x1, y1, x2, y2, ff.nrows, ff.ncols)
# Convert Cartesian line coordinate to CAMS compatible polar coordinates (flipped Y axis)
rho_cams, theta_cams = getPolarLine(x1, ff.nrows - y1, x2,
ff.nrows - y2, ff.nrows,
ff.ncols)
logDebug('converted rho, theta')
logDebug(rho, theta)
# Bounded the thresholded image by min and max frames
img = selectFrames(np.copy(img_thres), ff, frame_min, frame_max)
# Remove lonely pixels
img = morph.clean(img)
# Get indices of stripe pixels around the line
stripe_indices = getStripeIndices(rho, theta,
int(config.stripe_width * 1.5),
img.shape[0], img.shape[1])
# Extract the stripe from the thresholded image
stripe = np.zeros((ff.nrows, ff.ncols), np.uint8)
stripe[stripe_indices] = img[stripe_indices]
# Show detected line
# show('detected line: '+str(frame_min)+'-'+str(frame_max), stripe)
# Get stripe positions
stripe_positions = stripe.nonzero()
xs = stripe_positions[1]
ys = stripe_positions[0]
zs = ff.maxframe[stripe_positions]
# Make an array to feed into the centroiding algorithm
stripe_points = np.vstack((xs, ys, zs))
stripe_points = np.swapaxes(stripe_points, 0, 1)
# Sort stripe points by frame
stripe_points = stripe_points[stripe_points[:, 2].argsort()]
# Show 3D cloud
# show3DCloud(ff, stripe, detected_line, stripe_points, config)
# Get points of the given line
line_points = getAllPoints(stripe_points,
x1,
y1,
z1,
x2,
y2,
z2,
config,
fireball_detection=False)
# Skip if no points were returned
if not line_points.any():
continue
# Skip if the points cover too small a frame range
if abs(np.max(line_points[:, 2]) - np.min(line_points[:, 2])
) + 1 < config.line_minimum_frame_range_det:
continue
# Calculate centroids
centroids = []
for i in range(frame_min, frame_max + 1):
# Select pixel indicies belonging to a given frame
frame_pixels_inds = np.where(line_points[:, 2] == i)
# Get pixel positions in a given frame (pixels belonging to a found line)
frame_pixels = line_points[frame_pixels_inds].astype(np.int64)
# Get pixel positions in a given frame (pixels belonging to the whole stripe)
frame_pixels_stripe = stripe_points[np.where(
stripe_points[:, 2] == i)].astype(np.int64)
# Skip if there are no pixels in the frame
if not len(frame_pixels):
continue
# Calculate weights for centroiding
max_avg_corrected = ff.maxpixel - ff.avepixel
flattened_weights = (max_avg_corrected).astype(
np.float32) / ff.stdpixel
# Calculate centroids by half-frame
for half_frame in range(2):
# Apply deinterlacing if it is present in the video
if config.deinterlace_order >= 0:
# Deinterlace by fields (line lixels)
half_frame_pixels = frame_pixels[
frame_pixels[:, 1] %
2 == (config.deinterlace_order + half_frame) % 2]
# Deinterlace by fields (stripe pixels)
half_frame_pixels_stripe = frame_pixels_stripe[
frame_pixels_stripe[:, 1] %
2 == (config.deinterlace_order + half_frame) % 2]
# Skip if there are no pixels in the half-frame
if not len(half_frame_pixels):
continue
# Calculate half-frame value
frame_no = i + half_frame * 0.5
# No deinterlacing
else:
# Skip the second half frame
if half_frame == 1:
continue
half_frame_pixels = frame_pixels
half_frame_pixels_stripe = frame_pixels_stripe
frame_no = i
# Get maxpixel-avepixel values of given pixel indices (this will be used as weights)
max_weights = flattened_weights[half_frame_pixels[:, 1],
half_frame_pixels[:, 0]]
# Calculate weighted centroids
x_weighted = half_frame_pixels[:, 0] * np.transpose(
max_weights)
x_centroid = np.sum(x_weighted) / float(
np.sum(max_weights))
y_weighted = half_frame_pixels[:, 1] * np.transpose(
max_weights)
y_centroid = np.sum(y_weighted) / float(
np.sum(max_weights))
# Calculate intensity as the sum of threshold passer pixels on the stripe
#intensity_values = max_avg_corrected[half_frame_pixels[:,1], half_frame_pixels[:,0]]
intensity_values = max_avg_corrected[
half_frame_pixels_stripe[:, 1],
half_frame_pixels_stripe[:, 0]]
intensity = np.sum(intensity_values)
logDebug("centroid: ", frame_no, x_centroid, y_centroid,
intensity)
centroids.append(
[frame_no, x_centroid, y_centroid, intensity])
# Filter centroids
centroids = filterCentroids(centroids,
config.centroids_max_deviation,
config.centroids_max_distance)
# Convert to numpy array for easy slicing
centroids = np.array(centroids)
# Reject the solution if there are too few centroids
if len(centroids) < config.line_minimum_frame_range_det:
continue
# Check that the detection has a surface brightness above the background
# The assumption here is that the peak of the meteor should have the intensity which is at least
# that of a patch of 4x4 pixels that are of the mean background brightness
if np.max(centroids[:, 3]) < min_patch_intensity:
continue
# Check the detection if it has the proper angular velocity
if not checkAngularVelocity(centroids, config):
continue
# Append the result to the meteor detections
meteor_detections.append([rho_cams, theta_cams, centroids])
logDebug('time for processing:', time() - t_all)
# # Plot centroids to image
# fig, (ax1, ax2) = plt.subplots(nrows=2)
# ax1.imshow(ff.maxpixel - ff.avepixel, cmap='gray')
# ax1.scatter(centroids[:,1], centroids[:,2], s=5, c='r', edgecolors='none')
# # Plot lightcurve
# ax2.plot(centroids[:,0], centroids[:,3])
# # # Plot relative angular velocity
# # ang_vels = []
# # fr_prev, x_prev, y_prev, _ = centroids[0]
# # for fr, x, y, _ in centroids[1:]:
# # dx = x - x_prev
# # dy = y - y_prev
# # dfr = fr - fr_prev
# # ddist = np.sqrt(dx**2 + dy**2)
# # dt = dfr/config.fps
# # ang_vels.append(ddist/dt)
# # x_prev = x
# # y_prev = y
# # fr_prev = fr
# # ax2.plot(ang_vels)
# plt.show()
return meteor_detections