def get_bounding_box(keyword, url, filename):
# get the image
img = Image(url)
# resize the image so things aren't so slow, if necessary
w, h = img.size()
if w > 1200 or h > 1200:
maxdim = max(w, h)
ratio = math.ceil(maxdim/800.0)
print " resizing..."
img = img.resize(w=int(w/ratio), h=int(h/ratio))
else:
ratio = 1
# get the canvas
disp = Display((800, 800))
# text overlay
textlayer = DrawingLayer(img.size())
textlayer.setFontSize(30)
cx, cy = 10, 10
for xoff in range(-2, 3):
for yoff in range(-2, 3):
textlayer.text(keyword, (cx + xoff, cy + yoff), color=Color.BLACK)
textlayer.text(keyword, (cx, cy), color=Color.WHITE)
# two points to declare a bounding box
point1 = None
point2 = None
while disp.isNotDone():
cursor = (disp.mouseX, disp.mouseY)
if disp.leftButtonUp:
if point1 and point2:
point1 = None
point2 = None
if point1:
point2 = disp.leftButtonUpPosition()
else:
point1 = disp.leftButtonUpPosition()
bb = None
if point1 and point2:
bb = disp.pointsToBoundingBox(point1, point2)
elif point1 and not point2:
bb = disp.pointsToBoundingBox(point1, cursor)
img.clearLayers()
drawlayer = DrawingLayer(img.size())
if bb:
drawlayer.rectangle((bb[0], bb[1]), (bb[2], bb[3]), color=Color.RED)
# keyboard commands
if pygame.key.get_pressed()[pygame.K_s]:
# skip for now
raise Skip()
elif pygame.key.get_pressed()[pygame.K_b]:
# mark it as an invalid picture
raise BadImage()
elif pygame.key.get_pressed()[pygame.K_RETURN]:
if point1 and point2:
bb = disp.pointsToBoundingBox(scale(ratio, point1), scale(ratio, point2))
return bb
elif not point1 and not point2:
bb = disp.pointsToBoundingBox((0, 0), (w, h))
return bb
drawlayer.line((cursor[0], 0), (cursor[0], img.height), color=Color.BLUE)
drawlayer.line((0, cursor[1]), (img.width, cursor[1]), color=Color.BLUE)
#drawlayer.circle(cursor, 2, color=Color.BLUE, filled=True)
img.addDrawingLayer(textlayer)
img.addDrawingLayer(drawlayer)
img.save(disp)
def track(self):
print "Press right mouse button to pause or play"
print "Use left mouse button to select target"
print "Target color must be different from background"
print "Target must have width larger than height"
print "Target can be upside down"
#Parameters
isUDPConnection = False # Currently switched manually in the code
display = True
displayDebug = True
useBasemap = False
maxRelativeMotionPerFrame = 2 # How much the target can moved between two succesive frames
pixelPerRadians = 320
radius = pixelPerRadians
referenceImage = '../ObjectTracking/kite_detail.jpg'
scaleFactor = 0.5
isVirtualCamera = True
useHDF5 = False
# Open reference image: this is used at initlalisation
target_detail = Image(referenceImage)
# Get RGB color palette of target (was found to work better than using hue)
pal = target_detail.getPalette(bins = 2, hue = False)
# Open video to analyse or live stream
#cam = JpegStreamCamera('http://192.168.1.29:8080/videofeed')#640 * 480
if isVirtualCamera:
#cam = VirtualCamera('../../zenith-wind-power-read-only/KiteControl-Qt/videos/kiteFlying.avi','video')
#cam = VirtualCamera('/media/bat/DATA/Baptiste/Nautilab/kite_project/robokite/ObjectTracking/00095.MTS', 'video')
#cam = VirtualCamera('output.avi', 'video')
cam = VirtualCamera('../Recording/Videos/Flying kite images (for kite steering unit development)-YTMgX1bvrTo.flv','video')
virtualCameraFPS = 25
else:
cam = JpegStreamCamera('http://192.168.43.1:8080/videofeed')#640 * 480
#cam = Camera()
# Get a sample image to initialize the display at the same size
img = cam.getImage().scale(scaleFactor)
print img.width, img.height
# Create a pygame display
if display:
if img.width>img.height:
disp = Display((27*640/10,25*400/10))#(int(2*img.width/scaleFactor), int(2*img.height/scaleFactor)))
else:
disp = Display((810,1080))
#js = JpegStreamer()
# Initialize variables
previous_angle = 0 # target has to be upright when starting. Target width has to be larger than target heigth.
previous_coord_px = (0, 0) # Initialized to top left corner, which always exists
previous_dCoord = previous_coord_px
previous_dAngle = previous_angle
angles = []
coords_px = []
coord_px = [0, 0]
angle = 0
target_elevations = []
target_bearings = []
times = []
wasTargetFoundInPreviousFrame = False
i_frame = 0
isPaused = False
selectionInProgress = False
th = [100, 100, 100]
skycolor = Color.BLUE
timeLastTarget = 0
# Prepare recording
recordFilename = datetime.datetime.utcnow().strftime("%Y%m%d_%Hh%M_")+ 'simpleTrack'
if useHDF5:
try:
os.remove(recordFilename + '.hdf5')
except:
print('Creating file ' + recordFilename + '.hdf5')
""" The following line is used to silence the following error (according to http://stackoverflow.com/questions/15117128/h5py-in-memory-file-and-multiprocessing-error)
#000: ../../../src/H5F.c line 1526 in H5Fopen(): unable to open file
major: File accessability
minor: Unable to open file"""
h5py._errors.silence_errors()
recordFile = h5py.File(recordFilename + '.hdf5', 'a')
hdfSize = 0
dset = recordFile.create_dataset('kite', (2,2), maxshape=(None,7))
imset = recordFile.create_dataset('image', (2,img.width,img.height,3 ), maxshape=(None, img.width, img.height, 3))
else:
try:
os.remove(recordFilename + '.csv')
except:
print('Creating file ' + recordFilename + '.csv')
recordFile = file(recordFilename + '.csv', 'a')
csv_writer = csv.writer(recordFile)
csv_writer.writerow(['Time (s)', 'x (px)', 'y (px)', 'Orientation (rad)', 'Elevation (rad)', 'Bearing (rad)', 'ROT (rad/s)'])
# Launch a thread to get UDP message with orientation of the camera
mobile = mobileState.mobileState()
if isUDPConnection:
a = threading.Thread(None, mobileState.mobileState.checkUpdate, None, (mobile,))
a.start()
# Loop while not canceled by user
t0 = time.time()
previousTime = t0
while not(display) or disp.isNotDone():
t = time.time()
deltaT = (t-previousTime)
FPS = 1.0/deltaT
#print 'FPS =', FPS
if isVirtualCamera:
deltaT = 1.0/virtualCameraFPS
previousTime = t
i_frame = i_frame + 1
timestamp = datetime.datetime.utcnow()
# Receive orientation of the camera
if isUDPConnection:
mobile.computeRPY([2, 0, 1], [-1, 1, 1])
ctm = np.array([[sp.cos(mobile.roll), -sp.sin(mobile.roll)], \
[sp.sin(mobile.roll), sp.cos(mobile.roll)]]) # Coordinate transform matrix
if useBasemap:
# Warning this really slows down the computation
m = Basemap(width=img.width, height=img.height, projection='aeqd',
lat_0=sp.rad2deg(mobile.pitch), lon_0=sp.rad2deg(mobile.yaw), rsphere = radius)
# Get an image from camera
if not isPaused:
img = cam.getImage()
img = img.resize(int(scaleFactor*img.width), int(scaleFactor*img.height))
if display:
# Pause image when right button is pressed
dwn = disp.rightButtonDownPosition()
if dwn is not None:
isPaused = not(isPaused)
dwn = None
if display:
# Create a layer to enable user to make a selection of the target
selectionLayer = DrawingLayer((img.width, img.height))
if img:
if display:
# Create a new layer to host information retrieved from video
layer = DrawingLayer((img.width, img.height))
# Selection is a rectangle drawn while holding mouse left button down
if disp.leftButtonDown:
corner1 = (disp.mouseX, disp.mouseY)
selectionInProgress = True
if selectionInProgress:
corner2 = (disp.mouseX, disp.mouseY)
bb = disp.pointsToBoundingBox(corner1, corner2)# Display the temporary selection
if disp.leftButtonUp: # User has finished is selection
selectionInProgress = False
selection = img.crop(bb[0], bb[1], bb[2], bb[3])
if selection != None:
# The 3 main colors in the area selected are considered.
# Note that the selection should be included in the target and not contain background
try:
selection.save('../ObjectTracking/'+ 'kite_detail_tmp.jpg')
img0 = Image("kite_detail_tmp.jpg") # For unknown reason I have to reload the image...
pal = img0.getPalette(bins = 2, hue = False)
except: # getPalette is sometimes bugging and raising LinalgError because matrix not positive definite
pal = pal
wasTargetFoundInPreviousFrame = False
previous_coord_px = (bb[0] + bb[2]/2, bb[1] + bb[3]/2)
if corner1 != corner2:
selectionLayer.rectangle((bb[0], bb[1]), (bb[2], bb[3]), width = 5, color = Color.YELLOW)
# If the target was already found, we can save computation time by
# reducing the Region Of Interest around predicted position
if wasTargetFoundInPreviousFrame:
ROITopLeftCorner = (max(0, previous_coord_px[0]-maxRelativeMotionPerFrame/2*width), \
max(0, previous_coord_px[1] -height*maxRelativeMotionPerFrame/2))
ROI = img.crop(ROITopLeftCorner[0], ROITopLeftCorner[1], \
maxRelativeMotionPerFrame*width, maxRelativeMotionPerFrame*height, \
centered = False)
if display :
# Draw the rectangle corresponding to the ROI on the complete image
layer.rectangle((previous_coord_px[0]-maxRelativeMotionPerFrame/2*width, \
previous_coord_px[1]-maxRelativeMotionPerFrame/2*height), \
(maxRelativeMotionPerFrame*width, maxRelativeMotionPerFrame*height), \
color = Color.GREEN, width = 2)
else:
# Search on the whole image if no clue of where is the target
ROITopLeftCorner = (0, 0)
ROI = img
'''#Option 1
target_part0 = ROI.hueDistance(color=(142,50,65)).invert().threshold(150)
target_part1 = ROI.hueDistance(color=(93,16,28)).invert().threshold(150)
target_part2 = ROI.hueDistance(color=(223,135,170)).invert().threshold(150)
target_raw_img = target_part0+target_part1+target_part2
target_img = target_raw_img.erode(5).dilate(5)
#Option 2
target_img = ROI.hueDistance(imgModel.getPixel(10,10)).binarize().invert().erode(2).dilate(2)'''
# Find sky color
sky = (img-img.binarize()).findBlobs(minsize=10000)
if sky:
skycolor = sky[0].meanColor()
# Option 3
target_img = ROI-ROI # Black image
# Loop through palette of target colors
if display and displayDebug:
decomposition = []
i_col = 0
for col in pal:
c = tuple([int(col[i]) for i in range(0,3)])
# Search the target based on color
ROI.save('../ObjectTracking/'+ 'ROI_tmp.jpg')
img1 = Image('../ObjectTracking/'+ 'ROI_tmp.jpg')
filter_img = img1.colorDistance(color = c)
h = filter_img.histogram(numbins=256)
cs = np.cumsum(h)
thmax = np.argmin(abs(cs- 0.02*img.width*img.height)) # find the threshold to have 10% of the pixel in the expected color
thmin = np.argmin(abs(cs- 0.005*img.width*img.height)) # find the threshold to have 10% of the pixel in the expected color
if thmin==thmax:
newth = thmin
else:
newth = np.argmin(h[thmin:thmax]) + thmin
alpha = 0.5
th[i_col] = alpha*th[i_col]+(1-alpha)*newth
filter_img = filter_img.threshold(max(40,min(200,th[i_col]))).invert()
target_img = target_img + filter_img
#print th
i_col = i_col + 1
if display and displayDebug:
[R, G, B] = filter_img.splitChannels()
white = (R-R).invert()
r = R*1.0/255*c[0]
g = G*1.0/255*c[1]
b = B*1.0/255*c[2]
tmp = white.mergeChannels(r, g, b)
decomposition.append(tmp)
# Get a black background with with white target foreground
target_img = target_img.threshold(150)
target_img = target_img - ROI.colorDistance(color = skycolor).threshold(80).invert()
if display and displayDebug:
small_ini = target_img.resize(int(img.width/(len(pal)+1)), int(img.height/(len(pal)+1)))
for tmp in decomposition:
small_ini = small_ini.sideBySide(tmp.resize(int(img.width/(len(pal)+1)), int(img.height/(len(pal)+1))), side = 'bottom')
small_ini = small_ini.adaptiveScale((int(img.width), int(img.height)))
toDisplay = img.sideBySide(small_ini)
else:
toDisplay = img
#target_img = ROI.hueDistance(color = Color.RED).threshold(10).invert()
# Search for binary large objects representing potential target
target = target_img.findBlobs(minsize = 500)
if target: # If a target was found
if wasTargetFoundInPreviousFrame:
predictedTargetPosition = (width*maxRelativeMotionPerFrame/2, height*maxRelativeMotionPerFrame/2) # Target will most likely be close to the center of the ROI
else:
predictedTargetPosition = previous_coord_px
# If there are several targets in the image, take the one which is the closest of the predicted position
target = target.sortDistance(predictedTargetPosition)
# Get target coordinates according to minimal bounding rectangle or centroid.
coordMinRect = ROITopLeftCorner + np.array((target[0].minRectX(), target[0].minRectY()))
coord_px = ROITopLeftCorner + np.array(target[0].centroid())
# Rotate the coordinates of roll angle around the middle of the screen
rot_coord_px = np.dot(ctm, coord_px - np.array([img.width/2, img.height/2])) + np.array([img.width/2, img.height/2])
if useBasemap:
coord = sp.deg2rad(m(rot_coord_px[0], img.height-rot_coord_px[1], inverse = True))
else:
coord = localProjection(rot_coord_px[0]-img.width/2, img.height/2-rot_coord_px[1], radius, mobile.yaw, mobile.pitch, inverse = True)
target_bearing, target_elevation = coord
# Get minimum bounding rectangle for display purpose
minR = ROITopLeftCorner + np.array(target[0].minRect())
contours = target[0].contour()
contours = [ ROITopLeftCorner + np.array(contour) for contour in contours]
# Get target features
angle = sp.deg2rad(target[0].angle()) + mobile.roll
angle = sp.deg2rad(unwrap180(sp.rad2deg(angle), sp.rad2deg(previous_angle)))
width = target[0].width()
height = target[0].height()
# Check if the kite is upside down
# First rotate the kite
ctm2 = np.array([[sp.cos(-angle+mobile.roll), -sp.sin(-angle+mobile.roll)], \
[sp.sin(-angle+mobile.roll), sp.cos(-angle+mobile.roll)]]) # Coordinate transform matrix
rotated_contours = [np.dot(ctm2, contour-coordMinRect) for contour in contours]
y = [-tmp[1] for tmp in rotated_contours]
itop = np.argmax(y) # Then looks at the points at the top
ibottom = np.argmin(y) # and the point at the bottom
# The point the most excentered is at the bottom
if abs(rotated_contours[itop][0])>abs(rotated_contours[ibottom][0]):
isInverted = True
else:
isInverted = False
if isInverted:
angle = angle + sp.pi
# Filter the data
alpha = 1-sp.exp(-deltaT/self.filterTimeConstant)
if not(isPaused):
dCoord = np.array(previous_dCoord)*(1-alpha) + alpha*(np.array(coord_px) - previous_coord_px) # related to the speed only if cam is fixed
dAngle = np.array(previous_dAngle)*(1-alpha) + alpha*(np.array(angle) - previous_angle)
else :
dCoord = np.array([0, 0])
dAngle = np.array([0])
#print coord_px, angle, width, height, dCoord
# Record important data
times.append(timestamp)
coords_px.append(coord_px)
angles.append(angle)
target_elevations.append(target_elevation)
target_bearings.append(target_bearing)
# Export data to controller
self.elevation = target_elevation
self.bearing = target_bearing
self.orientation = angle
dt = time.time()-timeLastTarget
self.ROT = dAngle/dt
self.lastUpdateTime = t
# Save for initialisation of next step
previous_dCoord = dCoord
previous_angle = angle
previous_coord_px = (int(coord_px[0]), int(coord_px[1]))
wasTargetFoundInPreviousFrame = True
timeLastTarget = time.time()
else:
wasTargetFoundInPreviousFrame = False
if useHDF5:
hdfSize = hdfSize+1
dset.resize((hdfSize, 7))
imset.resize((hdfSize, img.width, img.height, 3))
dset[hdfSize-1,:] = [time.time(), coord_px[0], coord_px[1], angle, self.elevation, self.bearing, self.ROT]
imset[hdfSize-1,:,:,:] = img.getNumpy()
recordFile.flush()
else:
csv_writer.writerow([time.time(), coord_px[0], coord_px[1], angle, self.elevation, self.bearing, self.ROT])
if display :
if target:
# Add target features to layer
# Minimal rectange and its center in RED
layer.polygon(minR[(0, 1, 3, 2), :], color = Color.RED, width = 5)
layer.circle((int(coordMinRect[0]), int(coordMinRect[1])), 10, filled = True, color = Color.RED)
# Target contour and centroid in BLUE
layer.circle((int(coord_px[0]), int(coord_px[1])), 10, filled = True, color = Color.BLUE)
layer.polygon(contours, color = Color.BLUE, width = 5)
# Speed vector in BLACK
layer.line((int(coord_px[0]), int(coord_px[1])), (int(coord_px[0]+20*dCoord[0]), int(coord_px[1]+20*dCoord[1])), width = 3)
# Line giving angle
layer.line((int(coord_px[0]+200*sp.cos(angle)), int(coord_px[1]+200*sp.sin(angle))), (int(coord_px[0]-200*sp.cos(angle)), int(coord_px[1]-200*sp.sin(angle))), color = Color.RED)
# Line giving rate of turn
#layer.line((int(coord_px[0]+200*sp.cos(angle+dAngle*10)), int(coord_px[1]+200*sp.sin(angle+dAngle*10))), (int(coord_px[0]-200*sp.cos(angle + dAngle*10)), int(coord_px[1]-200*sp.sin(angle+dAngle*10))))
# Add the layer to the raw image
toDisplay.addDrawingLayer(layer)
toDisplay.addDrawingLayer(selectionLayer)
# Add time metadata
toDisplay.drawText(str(i_frame)+" "+ str(timestamp), x=0, y=0, fontsize=20)
# Add Line giving horizon
#layer.line((0, int(img.height/2 + mobile.pitch*pixelPerRadians)),(img.width, int(img.height/2 + mobile.pitch*pixelPerRadians)), width = 3, color = Color.RED)
# Plot parallels
for lat in range(-90, 90, 15):
r = range(0, 361, 10)
if useBasemap:
# \todo improve for high roll
l = m (r, [lat]*len(r))
pix = [np.array(l[0]), img.height-np.array(l[1])]
else:
l = localProjection(sp.deg2rad(r), \
sp.deg2rad([lat]*len(r)), \
radius, \
lon_0 = mobile.yaw, \
lat_0 = mobile.pitch, \
inverse = False)
l = np.dot(ctm, l)
pix = [np.array(l[0])+img.width/2, img.height/2-np.array(l[1])]
for i in range(len(r)-1):
if isPixelInImage((pix[0][i],pix[1][i]), img) or isPixelInImage((pix[0][i+1],pix[1][i+1]), img):
layer.line((pix[0][i],pix[1][i]), (pix[0][i+1], pix[1][i+1]), color=Color.WHITE, width = 2)
# Plot meridians
for lon in range(0, 360, 15):
r = range(-90, 91, 10)
if useBasemap:
# \todo improve for high roll
l = m ([lon]*len(r), r)
pix = [np.array(l[0]), img.height-np.array(l[1])]
else:
l= localProjection(sp.deg2rad([lon]*len(r)), \
sp.deg2rad(r), \
radius, \
lon_0 = mobile.yaw, \
lat_0 = mobile.pitch, \
inverse = False)
l = np.dot(ctm, l)
pix = [np.array(l[0])+img.width/2, img.height/2-np.array(l[1])]
for i in range(len(r)-1):
if isPixelInImage((pix[0][i],pix[1][i]), img) or isPixelInImage((pix[0][i+1],pix[1][i+1]), img):
layer.line((pix[0][i],pix[1][i]), (pix[0][i+1], pix[1][i+1]), color=Color.WHITE, width = 2)
# Text giving bearing
# \todo improve for high roll
for bearing_deg in range(0, 360, 30):
l = localProjection(sp.deg2rad(bearing_deg), sp.deg2rad(0), radius, lon_0 = mobile.yaw, lat_0 = mobile.pitch, inverse = False)
l = np.dot(ctm, l)
layer.text(str(bearing_deg), ( img.width/2+int(l[0]), img.height-20), color = Color.RED)
# Text giving elevation
# \todo improve for high roll
for elevation_deg in range(-60, 91, 30):
l = localProjection(0, sp.deg2rad(elevation_deg), radius, lon_0 = mobile.yaw, lat_0 = mobile.pitch, inverse = False)
l = np.dot(ctm, l)
layer.text(str(elevation_deg), ( img.width/2 ,img.height/2-int(l[1])), color = Color.RED)
#toDisplay.save(js)
toDisplay.save(disp)
if display :
toDisplay.removeDrawingLayer(1)
toDisplay.removeDrawingLayer(0)
recordFile.close()
def track(self):
print "Press right mouse button to pause or play"
print "Use left mouse button to select target"
print "Target color must be different from background"
print "Target must have width larger than height"
print "Target can be upside down"
#Parameters
isUDPConnection = False # Currently switched manually in the code
display = True
displayDebug = True
useBasemap = False
maxRelativeMotionPerFrame = 2 # How much the target can moved between two succesive frames
pixelPerRadians = 320
radius = pixelPerRadians
referenceImage = '../ObjectTracking/kite_detail.jpg'
scaleFactor = 0.5
isVirtualCamera = True
useHDF5 = False
# Open reference image: this is used at initlalisation
target_detail = Image(referenceImage)
# Get RGB color palette of target (was found to work better than using hue)
pal = target_detail.getPalette(bins=2, hue=False)
# Open video to analyse or live stream
#cam = JpegStreamCamera('http://192.168.1.29:8080/videofeed')#640 * 480
if isVirtualCamera:
#cam = VirtualCamera('../../zenith-wind-power-read-only/KiteControl-Qt/videos/kiteFlying.avi','video')
#cam = VirtualCamera('/media/bat/DATA/Baptiste/Nautilab/kite_project/robokite/ObjectTracking/00095.MTS', 'video')
#cam = VirtualCamera('output.avi', 'video')
cam = VirtualCamera(
'../Recording/Videos/Flying kite images (for kite steering unit development)-YTMgX1bvrTo.mp4',
'video')
virtualCameraFPS = 25
else:
cam = JpegStreamCamera(
'http://192.168.43.1:8080/videofeed') #640 * 480
#cam = Camera()
# Get a sample image to initialize the display at the same size
img = cam.getImage().scale(scaleFactor)
print img.width, img.height
# Create a pygame display
if display:
if img.width > img.height:
disp = Display(
(27 * 640 / 10, 25 * 400 / 10)
) #(int(2*img.width/scaleFactor), int(2*img.height/scaleFactor)))
else:
disp = Display((810, 1080))
#js = JpegStreamer()
# Initialize variables
previous_angle = 0 # target has to be upright when starting. Target width has to be larger than target heigth.
previous_coord_px = (
0, 0) # Initialized to top left corner, which always exists
previous_dCoord = previous_coord_px
previous_dAngle = previous_angle
angles = []
coords_px = []
coord_px = [0, 0]
angle = 0
target_elevations = []
target_bearings = []
times = []
wasTargetFoundInPreviousFrame = False
i_frame = 0
isPaused = False
selectionInProgress = False
th = [100, 100, 100]
skycolor = Color.BLUE
timeLastTarget = 0
# Prepare recording
recordFilename = datetime.datetime.utcnow().strftime(
"%Y%m%d_%Hh%M_") + 'simpleTrack'
if useHDF5:
try:
os.remove(recordFilename + '.hdf5')
except:
print('Creating file ' + recordFilename + '.hdf5')
""" The following line is used to silence the following error (according to http://stackoverflow.com/questions/15117128/h5py-in-memory-file-and-multiprocessing-error)
#000: ../../../src/H5F.c line 1526 in H5Fopen(): unable to open file
major: File accessability
minor: Unable to open file"""
h5py._errors.silence_errors()
recordFile = h5py.File(
os.path.join(os.getcwd(), 'log', recordFilename + '.hdf5'),
'a')
hdfSize = 0
dset = recordFile.create_dataset('kite', (2, 2),
maxshape=(None, 7))
imset = recordFile.create_dataset('image',
(2, img.width, img.height, 3),
maxshape=(None, img.width,
img.height, 3))
else:
try:
os.remove(recordFilename + '.csv')
except:
print('Creating file ' + recordFilename + '.csv')
recordFile = file(
os.path.join(os.getcwd(), 'log', recordFilename + '.csv'), 'a')
csv_writer = csv.writer(recordFile)
csv_writer.writerow([
'Time (s)', 'x (px)', 'y (px)', 'Orientation (rad)',
'Elevation (rad)', 'Bearing (rad)', 'ROT (rad/s)'
])
# Launch a thread to get UDP message with orientation of the camera
mobile = mobileState.mobileState()
if isUDPConnection:
mobile.open()
# Loop while not canceled by user
t0 = time.time()
previousTime = t0
while not (display) or disp.isNotDone():
t = time.time()
deltaT = (t - previousTime)
FPS = 1.0 / deltaT
#print 'FPS =', FPS
if isVirtualCamera:
deltaT = 1.0 / virtualCameraFPS
previousTime = t
i_frame = i_frame + 1
timestamp = datetime.datetime.utcnow()
# Receive orientation of the camera
if isUDPConnection:
mobile.computeRPY([2, 0, 1], [-1, 1, 1])
ctm = np.array([[sp.cos(mobile.roll), -sp.sin(mobile.roll)], \
[sp.sin(mobile.roll), sp.cos(mobile.roll)]]) # Coordinate transform matrix
if useBasemap:
# Warning this really slows down the computation
m = Basemap(width=img.width,
height=img.height,
projection='aeqd',
lat_0=sp.rad2deg(mobile.pitch),
lon_0=sp.rad2deg(mobile.yaw),
rsphere=radius)
# Get an image from camera
if not isPaused:
img = cam.getImage()
img = img.resize(int(scaleFactor * img.width),
int(scaleFactor * img.height))
if display:
# Pause image when right button is pressed
dwn = disp.rightButtonDownPosition()
if dwn is not None:
isPaused = not (isPaused)
dwn = None
if display:
# Create a layer to enable user to make a selection of the target
selectionLayer = DrawingLayer((img.width, img.height))
if img:
if display:
# Create a new layer to host information retrieved from video
layer = DrawingLayer((img.width, img.height))
# Selection is a rectangle drawn while holding mouse left button down
if disp.leftButtonDown:
corner1 = (disp.mouseX, disp.mouseY)
selectionInProgress = True
if selectionInProgress:
corner2 = (disp.mouseX, disp.mouseY)
bb = disp.pointsToBoundingBox(
corner1,
corner2) # Display the temporary selection
if disp.leftButtonUp: # User has finished is selection
selectionInProgress = False
selection = img.crop(bb[0], bb[1], bb[2], bb[3])
if selection != None:
# The 3 main colors in the area selected are considered.
# Note that the selection should be included in the target and not contain background
try:
selection.save('../ObjectTracking/' +
'kite_detail_tmp.jpg')
img0 = Image(
"kite_detail_tmp.jpg"
) # For unknown reason I have to reload the image...
pal = img0.getPalette(bins=2, hue=False)
except: # getPalette is sometimes bugging and raising LinalgError because matrix not positive definite
pal = pal
wasTargetFoundInPreviousFrame = False
previous_coord_px = (bb[0] + bb[2] / 2,
bb[1] + bb[3] / 2)
if corner1 != corner2:
selectionLayer.rectangle((bb[0], bb[1]),
(bb[2], bb[3]),
width=5,
color=Color.YELLOW)
# If the target was already found, we can save computation time by
# reducing the Region Of Interest around predicted position
if wasTargetFoundInPreviousFrame:
ROITopLeftCorner = (max(0, previous_coord_px[0]-maxRelativeMotionPerFrame/2*width), \
max(0, previous_coord_px[1] -height*maxRelativeMotionPerFrame/2))
ROI = img.crop(ROITopLeftCorner[0], ROITopLeftCorner[1], \
maxRelativeMotionPerFrame*width, maxRelativeMotionPerFrame*height, \
centered = False)
if display:
# Draw the rectangle corresponding to the ROI on the complete image
layer.rectangle((previous_coord_px[0]-maxRelativeMotionPerFrame/2*width, \
previous_coord_px[1]-maxRelativeMotionPerFrame/2*height), \
(maxRelativeMotionPerFrame*width, maxRelativeMotionPerFrame*height), \
color = Color.GREEN, width = 2)
else:
# Search on the whole image if no clue of where is the target
ROITopLeftCorner = (0, 0)
ROI = img
'''#Option 1
target_part0 = ROI.hueDistance(color=(142,50,65)).invert().threshold(150)
target_part1 = ROI.hueDistance(color=(93,16,28)).invert().threshold(150)
target_part2 = ROI.hueDistance(color=(223,135,170)).invert().threshold(150)
target_raw_img = target_part0+target_part1+target_part2
target_img = target_raw_img.erode(5).dilate(5)
#Option 2
target_img = ROI.hueDistance(imgModel.getPixel(10,10)).binarize().invert().erode(2).dilate(2)'''
# Find sky color
sky = (img - img.binarize()).findBlobs(minsize=10000)
if sky:
skycolor = sky[0].meanColor()
# Option 3
target_img = ROI - ROI # Black image
# Loop through palette of target colors
if display and displayDebug:
decomposition = []
i_col = 0
for col in pal:
c = tuple([int(col[i]) for i in range(0, 3)])
# Search the target based on color
ROI.save('../ObjectTracking/' + 'ROI_tmp.jpg')
img1 = Image('../ObjectTracking/' + 'ROI_tmp.jpg')
filter_img = img1.colorDistance(color=c)
h = filter_img.histogram(numbins=256)
cs = np.cumsum(h)
thmax = np.argmin(
abs(cs - 0.02 * img.width * img.height)
) # find the threshold to have 10% of the pixel in the expected color
thmin = np.argmin(
abs(cs - 0.005 * img.width * img.height)
) # find the threshold to have 10% of the pixel in the expected color
if thmin == thmax:
newth = thmin
else:
newth = np.argmin(h[thmin:thmax]) + thmin
alpha = 0.5
th[i_col] = alpha * th[i_col] + (1 - alpha) * newth
filter_img = filter_img.threshold(
max(40, min(200, th[i_col]))).invert()
target_img = target_img + filter_img
#print th
i_col = i_col + 1
if display and displayDebug:
[R, G, B] = filter_img.splitChannels()
white = (R - R).invert()
r = R * 1.0 / 255 * c[0]
g = G * 1.0 / 255 * c[1]
b = B * 1.0 / 255 * c[2]
tmp = white.mergeChannels(r, g, b)
decomposition.append(tmp)
# Get a black background with with white target foreground
target_img = target_img.threshold(150)
target_img = target_img - ROI.colorDistance(
color=skycolor).threshold(80).invert()
if display and displayDebug:
small_ini = target_img.resize(
int(img.width / (len(pal) + 1)),
int(img.height / (len(pal) + 1)))
for tmp in decomposition:
small_ini = small_ini.sideBySide(tmp.resize(
int(img.width / (len(pal) + 1)),
int(img.height / (len(pal) + 1))),
side='bottom')
small_ini = small_ini.adaptiveScale(
(int(img.width), int(img.height)))
toDisplay = img.sideBySide(small_ini)
else:
toDisplay = img
#target_img = ROI.hueDistance(color = Color.RED).threshold(10).invert()
# Search for binary large objects representing potential target
target = target_img.findBlobs(minsize=500)
if target: # If a target was found
if wasTargetFoundInPreviousFrame:
predictedTargetPosition = (
width * maxRelativeMotionPerFrame / 2,
height * maxRelativeMotionPerFrame / 2
) # Target will most likely be close to the center of the ROI
else:
predictedTargetPosition = previous_coord_px
# If there are several targets in the image, take the one which is the closest of the predicted position
target = target.sortDistance(predictedTargetPosition)
# Get target coordinates according to minimal bounding rectangle or centroid.
coordMinRect = ROITopLeftCorner + np.array(
(target[0].minRectX(), target[0].minRectY()))
coord_px = ROITopLeftCorner + np.array(
target[0].centroid())
# Rotate the coordinates of roll angle around the middle of the screen
rot_coord_px = np.dot(
ctm, coord_px -
np.array([img.width / 2, img.height / 2])) + np.array(
[img.width / 2, img.height / 2])
if useBasemap:
coord = sp.deg2rad(
m(rot_coord_px[0],
img.height - rot_coord_px[1],
inverse=True))
else:
coord = localProjection(
rot_coord_px[0] - img.width / 2,
img.height / 2 - rot_coord_px[1],
radius,
mobile.yaw,
mobile.pitch,
inverse=True)
target_bearing, target_elevation = coord
# Get minimum bounding rectangle for display purpose
minR = ROITopLeftCorner + np.array(target[0].minRect())
contours = target[0].contour()
contours = [
ROITopLeftCorner + np.array(contour)
for contour in contours
]
# Get target features
angle = sp.deg2rad(target[0].angle()) + mobile.roll
angle = sp.deg2rad(
unwrap180(sp.rad2deg(angle),
sp.rad2deg(previous_angle)))
width = target[0].width()
height = target[0].height()
# Check if the kite is upside down
# First rotate the kite
ctm2 = np.array([[sp.cos(-angle+mobile.roll), -sp.sin(-angle+mobile.roll)], \
[sp.sin(-angle+mobile.roll), sp.cos(-angle+mobile.roll)]]) # Coordinate transform matrix
rotated_contours = [
np.dot(ctm2, contour - coordMinRect)
for contour in contours
]
y = [-tmp[1] for tmp in rotated_contours]
itop = np.argmax(y) # Then looks at the points at the top
ibottom = np.argmin(y) # and the point at the bottom
# The point the most excentered is at the bottom
if abs(rotated_contours[itop][0]) > abs(
rotated_contours[ibottom][0]):
isInverted = True
else:
isInverted = False
if isInverted:
angle = angle + sp.pi
# Filter the data
alpha = 1 - sp.exp(-deltaT / self.filterTimeConstant)
if not (isPaused):
dCoord = np.array(previous_dCoord) * (
1 - alpha) + alpha * (
np.array(coord_px) - previous_coord_px
) # related to the speed only if cam is fixed
dAngle = np.array(previous_dAngle) * (
1 - alpha) + alpha * (np.array(angle) -
previous_angle)
else:
dCoord = np.array([0, 0])
dAngle = np.array([0])
#print coord_px, angle, width, height, dCoord
# Record important data
times.append(timestamp)
coords_px.append(coord_px)
angles.append(angle)
target_elevations.append(target_elevation)
target_bearings.append(target_bearing)
# Export data to controller
self.elevation = target_elevation
self.bearing = target_bearing
self.orientation = angle
dt = time.time() - timeLastTarget
self.ROT = dAngle / dt
self.lastUpdateTime = t
# Save for initialisation of next step
previous_dCoord = dCoord
previous_angle = angle
previous_coord_px = (int(coord_px[0]), int(coord_px[1]))
wasTargetFoundInPreviousFrame = True
timeLastTarget = time.time()
else:
wasTargetFoundInPreviousFrame = False
if useHDF5:
hdfSize = hdfSize + 1
dset.resize((hdfSize, 7))
imset.resize((hdfSize, img.width, img.height, 3))
dset[hdfSize - 1, :] = [
time.time(), coord_px[0], coord_px[1], angle,
self.elevation, self.bearing, self.ROT
]
imset[hdfSize - 1, :, :, :] = img.getNumpy()
recordFile.flush()
else:
csv_writer.writerow([
time.time(), coord_px[0], coord_px[1], angle,
self.elevation, self.bearing, self.ROT
])
if display:
if target:
# Add target features to layer
# Minimal rectange and its center in RED
layer.polygon(minR[(0, 1, 3, 2), :],
color=Color.RED,
width=5)
layer.circle(
(int(coordMinRect[0]), int(coordMinRect[1])),
10,
filled=True,
color=Color.RED)
# Target contour and centroid in BLUE
layer.circle((int(coord_px[0]), int(coord_px[1])),
10,
filled=True,
color=Color.BLUE)
layer.polygon(contours, color=Color.BLUE, width=5)
# Speed vector in BLACK
layer.line((int(coord_px[0]), int(coord_px[1])),
(int(coord_px[0] + 20 * dCoord[0]),
int(coord_px[1] + 20 * dCoord[1])),
width=3)
# Line giving angle
layer.line((int(coord_px[0] + 200 * sp.cos(angle)),
int(coord_px[1] + 200 * sp.sin(angle))),
(int(coord_px[0] - 200 * sp.cos(angle)),
int(coord_px[1] - 200 * sp.sin(angle))),
color=Color.RED)
# Line giving rate of turn
#layer.line((int(coord_px[0]+200*sp.cos(angle+dAngle*10)), int(coord_px[1]+200*sp.sin(angle+dAngle*10))), (int(coord_px[0]-200*sp.cos(angle + dAngle*10)), int(coord_px[1]-200*sp.sin(angle+dAngle*10))))
# Add the layer to the raw image
toDisplay.addDrawingLayer(layer)
toDisplay.addDrawingLayer(selectionLayer)
# Add time metadata
toDisplay.drawText(str(i_frame) + " " + str(timestamp),
x=0,
y=0,
fontsize=20)
# Add Line giving horizon
#layer.line((0, int(img.height/2 + mobile.pitch*pixelPerRadians)),(img.width, int(img.height/2 + mobile.pitch*pixelPerRadians)), width = 3, color = Color.RED)
# Plot parallels
for lat in range(-90, 90, 15):
r = range(0, 361, 10)
if useBasemap:
# \todo improve for high roll
l = m(r, [lat] * len(r))
pix = [np.array(l[0]), img.height - np.array(l[1])]
else:
l = localProjection(sp.deg2rad(r), \
sp.deg2rad([lat]*len(r)), \
radius, \
lon_0 = mobile.yaw, \
lat_0 = mobile.pitch, \
inverse = False)
l = np.dot(ctm, l)
pix = [
np.array(l[0]) + img.width / 2,
img.height / 2 - np.array(l[1])
]
for i in range(len(r) - 1):
if isPixelInImage(
(pix[0][i], pix[1][i]), img) or isPixelInImage(
(pix[0][i + 1], pix[1][i + 1]), img):
layer.line((pix[0][i], pix[1][i]),
(pix[0][i + 1], pix[1][i + 1]),
color=Color.WHITE,
width=2)
# Plot meridians
for lon in range(0, 360, 15):
r = range(-90, 91, 10)
if useBasemap:
# \todo improve for high roll
l = m([lon] * len(r), r)
pix = [np.array(l[0]), img.height - np.array(l[1])]
else:
l= localProjection(sp.deg2rad([lon]*len(r)), \
sp.deg2rad(r), \
radius, \
lon_0 = mobile.yaw, \
lat_0 = mobile.pitch, \
inverse = False)
l = np.dot(ctm, l)
pix = [
np.array(l[0]) + img.width / 2,
img.height / 2 - np.array(l[1])
]
for i in range(len(r) - 1):
if isPixelInImage(
(pix[0][i], pix[1][i]), img) or isPixelInImage(
(pix[0][i + 1], pix[1][i + 1]), img):
layer.line((pix[0][i], pix[1][i]),
(pix[0][i + 1], pix[1][i + 1]),
color=Color.WHITE,
width=2)
# Text giving bearing
# \todo improve for high roll
for bearing_deg in range(0, 360, 30):
l = localProjection(sp.deg2rad(bearing_deg),
sp.deg2rad(0),
radius,
lon_0=mobile.yaw,
lat_0=mobile.pitch,
inverse=False)
l = np.dot(ctm, l)
layer.text(
str(bearing_deg),
(img.width / 2 + int(l[0]), img.height - 20),
color=Color.RED)
# Text giving elevation
# \todo improve for high roll
for elevation_deg in range(-60, 91, 30):
l = localProjection(0,
sp.deg2rad(elevation_deg),
radius,
lon_0=mobile.yaw,
lat_0=mobile.pitch,
inverse=False)
l = np.dot(ctm, l)
layer.text(str(elevation_deg),
(img.width / 2, img.height / 2 - int(l[1])),
color=Color.RED)
#toDisplay.save(js)
toDisplay.save(disp)
if display:
toDisplay.removeDrawingLayer(1)
toDisplay.removeDrawingLayer(0)
recordFile.close()
from SimpleCV import Display, Camera, Image, DrawingLayer, VirtualCamera
disp = Display((600,800))
#cam = Camera()
cam = VirtualCamera('/media/bat/DATA/Baptiste/Nautilab/kite_project/zenith-wind-power-read-only/KiteControl-Qt/videos/kiteTest.avi','video')
isPaused = False
updateSelection = False
while(disp.isNotDone()):
if not isPaused:
img_flip = cam.getImage().flipHorizontal()
img = img_flip.edges(150, 100).dilate()
if disp.rightButtonDown:
isPaused = not(isPaused)
selectionLayer = DrawingLayer((img.width, img.height))
if disp.leftButtonDown:
corner1 = (disp.mouseX, disp.mouseY)
updateSelection = True
if updateSelection:
corner2 = (disp.mouseX, disp.mouseY)
bb = disp.pointsToBoundingBox(corner1, corner2)
if disp.leftButtonUp:
updateSelection = False
if corner1!=corner2:
selectionLayer.rectangle((bb[0],bb[1]),(bb[2],bb[3]))
img.addDrawingLayer(selectionLayer)
img.save(disp)
img.removeDrawingLayer(0)
from SimpleCV import Display, Camera, Image, DrawingLayer, VirtualCamera
disp = Display((600, 800))
#cam = Camera()
cam = VirtualCamera(
'../../Recording/Videos/kiteTest from Zenith Wind Power-fNMO3kAX0PE.mp4',
'video')
isPaused = False
updateSelection = False
while (disp.isNotDone()):
if not isPaused:
img_flip = cam.getImage().flipHorizontal()
img = img_flip.edges(150, 100).dilate()
if disp.rightButtonDown:
isPaused = not (isPaused)
selectionLayer = DrawingLayer((img.width, img.height))
if disp.leftButtonDown:
corner1 = (disp.mouseX, disp.mouseY)
updateSelection = True
if updateSelection:
corner2 = (disp.mouseX, disp.mouseY)
bb = disp.pointsToBoundingBox(corner1, corner2)
if disp.leftButtonUp:
updateSelection = False
if corner1 != corner2:
selectionLayer.rectangle((bb[0], bb[1]), (bb[2], bb[3]))
img.addDrawingLayer(selectionLayer)
img.save(disp)
img.removeDrawingLayer(0)