def svg_overlay(faces, frame_size, joy_score):
width, height = frame_size
doc = svg.Svg(width=width, height=height)
for face in faces:
x, y, w, h = face.bounding_box
doc.add(
svg.Rect(x=int(x),
y=int(y),
width=int(w),
height=int(h),
rx=10,
ry=10,
fill_opacity=0.3 * face.face_score,
style='fill:red;stroke:white;stroke-width:4px'))
doc.add(
svg.Text('Joy: %.2f' % face.joy_score,
x=x,
y=y - 10,
fill='red',
font_size=30))
doc.add(
svg.Text('Faces: %d Avg. joy: %.2f' % (len(faces), joy_score),
x=10,
y=50,
fill='red',
font_size=40))
return str(doc)
def plot_svg_chart(svg_doc, location, data):
x, y, x2, y2 = LOCATIONS[location]
plotted_data = sorted(list(data[location].keys()))[-20:]
for pos, time in enumerate(plotted_data):
start_x = (x + 8 * pos) * SVG_SCALE_FACTOR
width = 4 * SVG_SCALE_FACTOR
start_y = (y - 20 * data[location][time])
height = (y - start_y) * SVG_SCALE_FACTOR
start_y = start_y * SVG_SCALE_FACTOR
color = 'red' if data[location][time] == 2 else 'green'
svg_doc.add(
svg.Rect(x=int(start_x),
y=int(start_y),
width=int(width),
height=int(height),
fill_opacity=0.3,
style='fill:' + color + ';stroke:' + color +
';stroke-width:4px'))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--dog_park_model_path',
help='Path to the model file for the dog park.')
parser.add_argument('--vb1_model_path',
help='Path to the model file for volley ball court 1.')
parser.add_argument('--vb2_model_path',
help='Path to the model file for volley ball court 1.')
parser.add_argument(
'--label_path',
required=True,
help='Path to label file that corresponds to the model.')
parser.add_argument('--input_mean',
type=float,
default=128.0,
help='Input mean.')
parser.add_argument('--input_std',
type=float,
default=128.0,
help='Input std.')
parser.add_argument('--input_depth',
type=int,
default=3,
help='Input depth.')
parser.add_argument('--enable_streaming',
default=False,
action='store_true',
help='Enable streaming server')
parser.add_argument('--streaming_bitrate',
type=int,
default=1000000,
help='Streaming server video bitrate (kbps)')
parser.add_argument('--mdns_name',
default='',
help='Streaming server mDNS name')
parser.add_argument(
'--preview',
action='store_true',
default=False,
help=
'Enables camera preview in addition to printing result to terminal.')
parser.add_argument(
'--time_interval',
type=int,
default=10,
help='Time interval at which to store data in seconds.')
parser.add_argument(
'--gather_data',
action='store_true',
default=False,
help='Also save images according to the assigned category.')
parser.add_argument(
'--timelapse',
action='store_true',
default=False,
help='Also save some timelapses of the entire scene, every 120 seconds.'
)
parser.add_argument('--image_folder',
default='/home/pi/Pictures/Data',
help='Folder to save captured images')
args = parser.parse_args()
labels = read_labels(args.label_path)
# At least one model needs to be passed in.
assert args.dog_park_model_path or args.vb1_model_path or args.vb2_model_path
# Check that the folder exists
if args.gather_data:
expected_subfolders = ['dog_park', 'court_one', 'court_two']
subfolders = os.listdir(args.image_folder)
for folder in expected_subfolders:
assert folder in subfolders
with ExitStack() as stack:
dog_park = {
'location_name': 'dog_park',
'path': args.dog_park_model_path,
} if args.dog_park_model_path else None
vb1 = {
'location_name': 'court_one',
'path': args.vb1_model_path,
} if args.vb1_model_path else None
vb2 = {
'location_name': 'court_two',
'path': args.vb2_model_path,
} if args.vb2_model_path else None
# Get the list of models, filter to only the ones that were passed in.
models = [dog_park, vb1, vb2]
models = list(filter(lambda model: model, models))
# Initialize models and add them to the context
for model in models:
print('Initializing {model_name}...'.format(
model_name=model["location_name"]))
descriptor = inference.ModelDescriptor(
name='mobilenet_based_classifier',
input_shape=(1, 160, 160, args.input_depth),
input_normalizer=(args.input_mean, args.input_std),
compute_graph=utils.load_compute_graph(model['path']))
model['descriptor'] = descriptor
if dog_park:
dog_park['image_inference'] = stack.enter_context(
inference.ImageInference(dog_park['descriptor']))
if vb1:
vb1['image_inference'] = stack.enter_context(
inference.ImageInference(vb1['descriptor']))
if vb2:
vb2['image_inference'] = stack.enter_context(
inference.ImageInference(vb2['descriptor']))
camera = stack.enter_context(
PiCamera(sensor_mode=4, resolution=(820, 616), framerate=30))
server = None
if args.enable_streaming:
server = stack.enter_context(
StreamingServer(camera,
bitrate=args.streaming_bitrate,
mdns_name=args.mdns_name))
if args.preview:
# Draw bounding boxes around locations
# Load the arbitrarily sized image
img = Image.new('RGB', (820, 616))
draw = ImageDraw.Draw(img)
for location in LOCATIONS.values():
x1, y1, x2, y2 = location
draw_rectangle(draw, x1, y1, x2, y2, 3, outline='white')
# Create an image padded to the required size with
# mode 'RGB'
pad = Image.new('RGB', (
((img.size[0] + 31) // 32) * 32,
((img.size[1] + 15) // 16) * 16,
))
# Paste the original image into the padded one
pad.paste(img, (0, 0))
# Add the overlay with the padded image as the source,
# but the original image's dimensions
camera.add_overlay(pad.tobytes(), alpha=64, layer=3, size=img.size)
camera.start_preview()
data_filename = _make_filename(args.image_folder, 'data', None, 'json')
data_generator = commit_data_to_long_term(args.time_interval,
data_filename)
data_generator.send(None)
# Capture one picture of entire scene each time it's started again.
time.sleep(2)
date = time.strftime('%Y-%m-%d')
scene_filename = _make_filename(args.image_folder, date, None)
camera.capture(scene_filename)
# Draw bounding box on image showing the crop locations
with Image.open(scene_filename) as scene:
draw = ImageDraw.Draw(scene)
for location in LOCATIONS.values():
x1, y1, x2, y2 = location
draw_rectangle(draw, x1, y1, x2, y2, 3, outline='white')
scene.save(scene_filename)
# Constantly get cropped images
for cropped_images in get_cropped_images(camera, args.timelapse):
svg_doc = None
if args.enable_streaming:
width = 820 * SVG_SCALE_FACTOR
height = 616 * SVG_SCALE_FACTOR
svg_doc = svg.Svg(width=width, height=height)
for location in LOCATIONS.values():
x, y, x2, y2 = location
w = (x2 - x) * SVG_SCALE_FACTOR
h = (y2 - y) * SVG_SCALE_FACTOR
x = x * SVG_SCALE_FACTOR
y = y * SVG_SCALE_FACTOR
svg_doc.add(
svg.Rect(
x=int(x),
y=int(y),
width=int(w),
height=int(h),
rx=10,
ry=10,
fill_opacity=0.3,
style='fill:none;stroke:white;stroke-width:4px'))
# For each inference model, crop and process a different thing.
for model in models:
location_name = model['location_name']
image_inference = model['image_inference']
cropped_image = cropped_images[location_name]
# TODO: (Image Comparison) If False,return no activity.
if cropped_image:
# then run image_inference on them.
result = image_inference.run(cropped_image)
processed_result = process(result, labels, 'final_result')
data_generator.send(
(location_name, processed_result, svg_doc))
message = get_message(processed_result)
# Print the message
# print('\n')
# print('{location_name}:'.format(location_name=location_name))
# print(message)
else:
# Fake processed_result
processed_result = [('inactive', 1.00), ('active', 0.00)]
data_generator.send(
(location_name, processed_result, svg_doc))
label = processed_result[0][0]
timestamp = time.strftime('%Y-%m-%d_%H.%M.%S')
# print(timestamp)
# print('\n')
if args.gather_data and cropped_image:
# Gather 1% data on 'no activity' since it's biased against that.
# Gather 0.1% of all images.
if (
# (label == 'no activity' and random.random() > 0.99) or
# (random.random() > 0.999)
# (location_name != 'dog_park' and random.random() > 0.99) or
(random.random() > 0.9)):
subdir = '{location_name}/{label}'.format(
location_name=location_name, label=label)
filename = _make_filename(args.image_folder, timestamp,
subdir)
cropped_image.save(filename)
# if svg_doc:
# ## Plot points out
# ## 160 x 80 grid
# ## 16px width
# ## 20, 40, 60 for 0, 1, 2
# lines = message.split('\n')
# y_correction = len(lines) * 20
# for line in lines:
# svg_doc.add(svg.Text(line,
# x=(LOCATIONS[location_name][0]) * SVG_SCALE_FACTOR,
# y=(LOCATIONS[location_name][1] - y_correction) * SVG_SCALE_FACTOR,
# fill='white', font_size=20))
# y_correction = y_correction - 20
# TODO: Figure out how to annotate at specific locations.
# if args.preview:
# camera.annotate_foreground = Color('black')
# camera.annotate_background = Color('white')
# # PiCamera text annotation only supports ascii.
# camera.annotate_text = '\n %s' % message.encode(
# 'ascii', 'backslashreplace').decode('ascii')
if server:
server.send_overlay(str(svg_doc))
if args.preview:
camera.stop_preview()
