def show_tracks(scene, tracks, frame_dets=(), first_frame=None):
if not tracks:
return
if first_frame is None:
first_frame = min(tr[0].frame for tr in tracks)
last_frame = max(tr[-1].frame for tr in tracks)
for f in range(first_frame, last_frame + 1):
img = scene.frame(f)
if f in frame_dets:
for d in frame_dets[f]:
d.draw(img, color=(255, 0, 0))
for tr_id, tr in enumerate(tracks):
if tr[0].frame <= f <= tr[-1].frame:
det = tr[bisect([det.frame for det in tr], f) - 1]
assert det.frame == f # Did you interpolate_missing_detections(tracks)?
if det.id is not None:
label = '%s:%s' % (tr_id, det.id)
else:
label = tr_id
det.draw(img, label=label)
cv2.polylines(img,
np.array([scene.roi()]),
True, (0, 0, 0),
thickness=3)
cv2.polylines(img,
np.array([scene.roi()]),
True, (255, 255, 255),
thickness=1)
view(img) #, pause=True)
imwrite(img, "dbg/%.8d.jpg" % f)
def make_graph(video_detections, fps, show=False):
max_temporal_distance = fps
min_iou = 0.0
history = []
graph = []
for frame_idx, frame, detections in video_detections:
for det in detections:
det.next_weight_data = defaultdict(list)
det.prev = set()
graph.append(det)
if show:
det.draw(frame, label=det.id)
history = history[-max_temporal_distance:]
for old_detections in history:
for prv in old_detections:
for nxt in detections:
if nxt.iou(prv) > min_iou:
connect(prv, nxt, None)
history.append(detections)
if show:
view(frame, pause=True)
return graph
def test_flipp_aspect_ratio_single_image_wide():
"""Test flipp with only one image wider than 16:9
"""
from vi3o import view, flipp
from vi3o.debugview import DebugViewer
img_1 = np.zeros(shape=(5, 10), dtype=np.uint8)
flipp()
view(img_1)
flipp()
assert DebugViewer.named_viewers["Default"].image.width == 10
assert DebugViewer.named_viewers["Default"].image.height == 5
def test_flipp_aspect_ratio_default_viewer_1_1_vertical():
"""Test exactly 1:1 aspect ratio stacked horizontally
"""
from vi3o import view, flipp
from vi3o.debugview import DebugViewer
img_1 = np.zeros(shape=(3, 9), dtype=np.uint8)
img_2 = np.ones(shape=(3, 9), dtype=np.uint8) * 128
img_3 = np.ones(shape=(3, 9), dtype=np.uint8) * 255
flipp()
view(img_1)
view(img_2)
view(img_3)
flipp(aspect_ratio=1)
assert DebugViewer.named_viewers["Default"].image.width == 9
assert DebugViewer.named_viewers["Default"].image.height == 9
def test_flipp_aspect_ratio_default_viewer_default_is_16_9_approx_narrow():
"""Test approimatelly 16:9 aspect ratio
"""
from vi3o import view, flipp
from vi3o.debugview import DebugViewer
img_1 = np.zeros(shape=(9, 5), dtype=np.uint8)
img_2 = np.ones(shape=(9, 5), dtype=np.uint8) * 128
img_3 = np.ones(shape=(9, 5), dtype=np.uint8) * 255
flipp()
view(img_1)
view(img_2)
view(img_3)
flipp()
assert DebugViewer.named_viewers["Default"].image.width == 15
assert DebugViewer.named_viewers["Default"].image.height == 9
def test_flipp_aspect_ratio_default_viewer_infinite_vertical():
"""Test infinite horizontal stacking, this is the legacy default
"""
from vi3o import view, flipp
from vi3o.debugview import DebugViewer
img_1 = np.zeros(shape=(10, 1), dtype=np.uint8)
img_2 = np.ones(shape=(10, 1), dtype=np.uint8) * 128
img_3 = np.ones(shape=(10, 1), dtype=np.uint8) * 255
flipp()
view(img_1)
view(img_2)
view(img_3)
flipp(aspect_ratio=0)
assert DebugViewer.named_viewers["Default"].image.width == 1
assert DebugViewer.named_viewers["Default"].image.height == 30
def imshowsc(img):
"""
Rescales (and translates) the intensities of the image *img* to cover the 0..255 range.
Then display the image *img* in the DebugViewer and pause the viewer with the image showing.
"""
view(img, scale=True, pause=True)
def imshow(img):
"""
Display the image *img* in the DebugViewer and pause the viewer with the image showing.
"""
view(img, pause=True)
def __next__(self):
return self.next()
def next(self):
headers = {}
while True:
l = self._fd.readline()
if not l:
raise StopIteration
if not l.strip() and headers:
break
if b':' in l:
i = l.index(b':')
headers[l[:i]] = l[i + 1:].strip()
data = self._fd.read(int(headers[b'Content-Length']))
img = imread(StringIO(data)).view(Frame)
img.index = self.fcnt
self.fcnt += 1
img.timestamp = img.systime = -1 # FIXME
return img
if __name__ == '__main__':
from vi3o import view
for img in AxisCam("192.168.0.90",
username="******",
password="******",
no_proxy=True):
view(img)
request.add_header("Authorization", "Basic %s" % base64string)
self._fd = urllib2.urlopen(request)
self.fcnt = 0
def __iter__(self):
return self
def next(self):
headers = {}
while True:
l = self._fd.readline()
if not l:
raise StopIteration
if not l.strip() and headers:
break
if ':' in l:
i = l.index(':')
headers[l[:i]] = l[i+1:].strip()
data = self._fd.read(int(headers['Content-Length']))
img = imread(StringIO(data)).view(Frame)
img.index = self.fcnt
self.fcnt += 1
img.timestamp = img.systime = -1 # FIXME
return img
if __name__ == '__main__':
from vi3o import view
for img in AxisCam('10.2.3.125', 640, 360):
view(img)
def imshowsc(img):
"""
Rescales (and translates) the intensities of the image *img* to cover the 0..255 range.
Then display the image *img* in the DebugViewer and pause the viewer with the image showing.
"""
view(img, scale=True, pause=True)
def imshow(img):
"""
Display the image *img* in the DebugViewer and pause the viewer with the image showing.
"""
view(img, pause=True)
def make_graph(video_detections, fps, show=False, max_connect=5):
tracks = []
lk_params = dict(winSize=(15, 15),
maxLevel=4,
criteria=(cv2.TERM_CRITERIA_EPS | cv2.TERM_CRITERIA_COUNT,
10, 0.03))
feature_params = dict(maxCorners=5000,
qualityLevel=0.01,
minDistance=10,
blockSize=7)
col = (255, 0, 0)
max_len = 3 * fps
min_klt_per_obj = 10
velocity_history = fps // 2
prediction_df = fps
graph = []
detect = True
for frame_idx, frame, detections in video_detections:
frame_gray = cv2.cvtColor(frame, cv2.COLOR_RGB2GRAY)
height, width, _ = frame.shape
estimate_intradet_iou(detections)
for det in detections:
det.next_weight_data = defaultdict(list)
det.pre_vs = []
det.post_vs = []
det.prev = set()
if show:
det.draw(frame, label=det.id)
graph.append(det)
# Track klt points to next frame
if len(tracks) > 0:
interesting = []
img0, img1 = prev_gray, frame_gray
p0 = np.float32([(tr.x, tr.y) for tr in tracks]).reshape(-1, 1, 2)
# See how the points have moved between the two frames
p1, st, err1 = cv2.calcOpticalFlowPyrLK(img0, img1, p0, None,
**lk_params)
p0r, st, err = cv2.calcOpticalFlowPyrLK(img1, img0, p1, None,
**lk_params)
d = abs(p0 - p0r).reshape(-1, 2).max(-1)
good = d < 1
new_tracks = []
for tr, (x, y), good_flag, e in zip(tracks, p1.reshape(-1, 2),
good, err1.flat):
if not good_flag:
continue
if not (0 <= x < width and 0 <= y < height):
continue
if e > 1e3:
continue
tr.history.append((frame_idx, x, y, e))
tr.history = tr.history[-max_len - 1:]
new_tracks.append(tr)
if show:
cv2.circle(frame, (x, y), 2,
(255 - min(e * 10, 255), 0, 0), -1)
for prev_dets in tr.dets_history:
for det in prev_dets:
if det.id == 2860144:
interesting.append(tr)
tracks = new_tracks
interesting = tracks
if show:
cv2.polylines(frame, [
np.int32([(x, y) for _, x, y, _ in tr.history])
for tr in interesting
], False, col)
# Find detections with too few klt points
if detect:
mask = np.zeros_like(frame_gray)
detect = False
min_area = float('Inf')
for det in detections:
cnt = 0
for tr in tracks:
if det.covers(tr.x, tr.y):
cnt += 1
if cnt < min_klt_per_obj:
det.update_mask(mask)
detect = True
min_area = min(min_area, det.area)
# Detect new klt points
if detect:
feature_params['minDistance'] = int(
np.sqrt(min_area / min_klt_per_obj))
for tr in tracks:
cv2.circle(mask, (tr.x, tr.y),
feature_params['minDistance'] // 2, 0, -1)
p = cv2.goodFeaturesToTrack(frame_gray,
mask=mask,
**feature_params)
if p is not None:
for x, y in np.float32(p).reshape(-1, 2):
nt = KltTrack(frame_idx, x, y)
tracks.append(nt)
# Assign detections to klt points and build detection connectivity graph
new_tracks = []
for tr in tracks:
vx = vy = None
last_dets = []
for nxt in detections:
if nxt.covers(tr.x, tr.y):
last_dets.append(nxt)
tr.dets_history_for_post_vx[frame_idx].append(nxt)
for prev_dets in tr.dets_history:
for prv in prev_dets:
df = nxt.frame - prv.frame
klt = tr.history[-df - 1:]
connect(prv, nxt, ('klt', klt))
if vx is None and len(
tr.history
) > velocity_history: # Predic where the detection will be in the future
hist = tr.history[-velocity_history:]
(vx,
x0), rx, _, _, _ = np.polyfit(range(len(hist)),
[p[1] for p in hist],
1,
full=True)
(vy,
y0), ry, _, _, _ = np.polyfit(range(len(hist)),
[p[2] for p in hist],
1,
full=True)
klt_res = [p[3] for p in hist]
r = (rx[0] + ry[0], sum(klt_res), max(klt_res))
if vx is not None:
v = (vx, vy) + r
nxt.pre_vs.append(v)
for df in range(1, prediction_df):
d = nxt.predict(df, vx, vy)
d.original = nxt
d.prediction_v = v
tr.predictions[d.frame].append(d)
for d in tr.dets_history_for_post_vx[frame_idx -
velocity_history]:
d.post_vs.append(v)
if frame_idx - velocity_history in tr.dets_history_for_post_vx:
del tr.dets_history_for_post_vx[frame_idx - velocity_history]
if last_dets:
tr.dets_history.append(last_dets)
tr.dets_history = tr.dets_history[-max_connect:]
f = frame_idx - max_len
tr.dets_history = [
last_dets for last_dets in tr.dets_history
if last_dets[0].frame > f
]
if tr.dets_history:
new_tracks.append(tr)
tracks = new_tracks
# Form long term connection from predicted detections
for tr in tracks:
for prd in tr.predictions[frame_idx]:
for det in detections:
if det not in tr.dets_history[-1] and prd.iou(det) > 0.5:
connect(prd.original, det, ('long', prd))
del tr.predictions[frame_idx]
if show:
for det in detections:
if det.pre_vs:
# vx = np.median([vx for vx, vy in det.pre_vs])
# vy = np.median([vy for vx, vy in det.pre_vs])
for vx, vy, r, _, _ in det.pre_vs:
df = 30
d = det.predict(df, vx, vy)
cv2.arrowedLine(frame, (int(det.cx), int(det.cy)),
(int(d.cx), int(d.cy)),
(0, 0, max(0, 255 - int(r))), 1)
prev_gray = frame_gray
if show:
view(frame)
return graph
