def get_ride_heading(ride, variations=False, moving_average_window=3, stops=False, version=1):
'''
I don't know exactly what this does. I was drunk when I wrote it.
'''
ride = np.array(ride)
if moving_average_window:
ride = util.movingaverage(ride, moving_average_window)
ROLL_STEP = 3
ride2 = np.array(ride)
ride1 = np.roll(ride2, ROLL_STEP, axis=0)
l = len(ride)
ride0 = np.hstack((np.ones((l,1)), np.zeros((l,1))))
ride0 = ride0[ROLL_STEP:]
ride1 = ride1[ROLL_STEP:]
ride2 = ride2[ROLL_STEP:]
a1 = ride0
a2 = ride2 - ride1
distances1 = np.linalg.norm(a1, axis=1)
distances = np.linalg.norm(a2, axis=1)
x = (a1 * a2).sum(1) / np.maximum(distances1 * distances, 0.1)
y = np.sign(a2[:,1])
np.seterr(all='ignore')
angles = np.arccos(x) * 180 / np.pi
np.seterr(all='print')
angles[y<0] = 360 - angles[y<0]
angles[distances < 2] = np.nan
is_stopped = []
angle = []
angle_window = []
WINDOW_SIZE = 2
MIN_SPEED = 2
direction = []
for i, dist in enumerate(distances):
if dist > MIN_SPEED:
angle_window.append(a2[i])
angle_window = angle_window[-WINDOW_SIZE:]
if len(angle_window) < WINDOW_SIZE:
direction.append(np.nan)
else:
d = get_angle(angle_window[-WINDOW_SIZE/2:]) - get_angle(angle_window[:WINDOW_SIZE/2])
if d > 180:
d -= 360
if d < -180:
d += 360
direction.append(d)
else:
direction.append(np.nan)
if dist < MIN_SPEED or len(angle_window) < WINDOW_SIZE:
is_stopped.append(True)
else:
is_stopped.append(False)
angle.append(get_angle(angle_window))
windows = []
current_window = []
current_window_type = 0
for i in range(len(direction)):
if np.isnan(direction[i]):
current_window = []
windows.append(0)
continue
current_window.append(direction[i])
current_window = current_window[-4:]
t = np.mean(current_window)
if current_window_type == 0:
if np.abs(t) > 3:
current_window_type = np.sign(t)
else:
if np.sign(current_window[-1]) != current_window_type:
current_window_type = 0
windows.append(current_window_type)
windows = windows[2:] + [0, 0]
sw = True
while sw:
sw = False
for i in range(1, len(windows) - 1):
if windows[i] != windows[i-1] and windows[i] != windows[i+1]:
windows[i] = windows[i+1]
sw = True
for i in range(3, len(windows)):
if windows[i-3] != windows[i-2] and \
windows[i-2] == windows[i-1] and \
windows[i-1] != windows[i]:
windows[i-2] = windows[i-3]
windows[i-1] = windows[i]
sw = True
description = []
current_window_type = 'stop' if stops else 'straight'
new_type = current_window_type
current_window_length = 0
for i in range(1, len(windows)):
if stops:
if is_stopped[i]:
if current_window_type != 'stop':
new_type = 'stop'
if windows[i] == 0 and not is_stopped[i]:
if current_window_type != 'straight':
new_type = 'straight'
else:
if windows[i] == 0 or is_stopped[i]:
if current_window_type != 'straight':
new_type = 'straight'
if windows[i] == 1:
if current_window_type != 'left':
new_type = 'left'
if windows[i] == -1:
if current_window_type != 'right':
new_type = 'right'
if new_type == current_window_type:
current_window_length += 1
else:
if current_window_length:
description.append([
current_window_type,
current_window_length,
util.euclidian_distance(ride2[i-current_window_length], ride2[i]),
-np.sum(direction[i-current_window_length : i-1])
])
current_window_type = new_type
current_window_length = 0
i = 0
while i < len(description) - 1:
if description[i][0] in ['right', 'left'] and np.abs(description[i][3]) < 15:
description[i+1][2] += description[i][2]
description.pop(i)
else:
i += 1
i = 0
while i < len(description) - 1:
if description[i][0] == description[i+1][0]:
description[i+1][1] += description[i][1]
description[i+1][2] += description[i][2]
description[i+1][3] += description[i][3]
description.pop(i)
else:
i += 1
# convert to words
DIST_TH = [0, 10, 50, 100, 250, 500, 1500]
if version == 1:
TIME_TH = [0, 2, 8, 50]
ANGLE_TH = [0, 35, 70, 110, 145]
else:
TIME_TH = [0, 8, 50]
ANGLE_TH = [0, 10, 30, 55, 80, 110, 145]
mirrored = {'straight': 'straight', 'left': 'right', 'right': 'left', 'stop': 'stop'}
words_original = []
words_mirror = []
for row in description:
if row[0] == 'stop':
v = np.digitize([row[1]], TIME_TH)[0]
elif row[0] == 'straight':
v = np.digitize([row[2]], DIST_TH)[0]
else:
v = np.digitize([np.abs(row[3])], ANGLE_TH)[0]
word = '%s_%s' % (row[0], v)
words_original.append(word)
word = '%s_%s' % (mirrored[row[0]], v)
words_mirror.append(word)
if variations:
words_inverted = list(reversed(words_mirror))
words_mirror_inverted = list(reversed(words_original))
return [words_original, words_mirror, words_inverted, words_mirror_inverted]
return words_original
def annotate_events(self, df, index_col='frameIndex'):
self.stats_file = open('%s/stats.txt' % self.data_direc, 'a')
self.stats_file.seek(0)
self.stats_file.truncate()
df['gz'] = util.movingaverage(df['gz'], self.moving_average_size)
windowed_df_test = util.generate_windows(df, window=self.window_size, ignore_columns=self.ignore_columns)
windowed_df_test = windowed_df_test[self.active_features]
predicted_labels_test = self.model.predict(windowed_df_test)
events = {
"left_lc_start": set(),
"left_lc_end" : set(),
"right_lc_start": set(),
"right_lc_end" : set(),
}
left_lc_start = 0
right_lc_start = 0
left_lc_end = 0
right_lc_end = 0
pos_label = 2
neg_label = 1
null_label = 0
l_start = OrderedDict()
r_start = OrderedDict()
for i in xrange(len(predicted_labels_test) - 4):
# starts with OO
if predicted_labels_test[i+1] == pos_label and len(set(predicted_labels_test[i+1:i+4])) == 1 \
and (predicted_labels_test[i] == null_label or predicted_labels_test[i] == neg_label or i == 0):
l_start[i] = 0
for k in r_start.keys():
r_start[k] += 1
# starts with <<
if predicted_labels_test[i+1] == neg_label and len(set(predicted_labels_test[i+1:i+4])) == 1 \
and (predicted_labels_test[i] == null_label or predicted_labels_test[i] == pos_label or i == 0):
r_start[i] = 0
for k in l_start.keys():
l_start[k] += 1
# ends with OO
if predicted_labels_test[i] == pos_label and predicted_labels_test[i+1] == pos_label \
and (predicted_labels_test[i+2] == null_label or predicted_labels_test[i+2] == neg_label):
found = False
for k, v in l_start.items():
if v >= 1:
del l_start[k]
if found:
continue
signal = df.iloc[k:i]['theta'].tolist()
if self.is_valid_event(signal, k, i, 'left'):
if (len(events['right_lc_end']) > 0 and k > max(events['right_lc_end']) or len(events['right_lc_end']) == 0) \
and (len(events['left_lc_end']) > 0 and k > max(events['left_lc_end']) or len(events['left_lc_end']) == 0):
events['left_lc_start'].add(k)
events['left_lc_end'].add(i)
found = True
# ends with <<
if predicted_labels_test[i] == neg_label and predicted_labels_test[i+1] == neg_label \
and (predicted_labels_test[i+2] == null_label or predicted_labels_test[i+2] == pos_label):
found = False
for k, v in r_start.items():
if v >= 1:
del r_start[k]
if found:
continue
signal = df.iloc[k:i]['theta'].tolist()
if self.is_valid_event(signal, k, i, 'right'):
if (len(events['right_lc_end']) > 0 and k > max(events['right_lc_end']) or len(events['right_lc_end']) == 0) \
and (len(events['left_lc_end']) > 0 and k > max(events['left_lc_end']) or len(events['left_lc_end']) == 0):
events['right_lc_start'].add(k)
events['right_lc_end'].add(i)
found = True
for k, v in events.iteritems():
events[k] = sorted(list(v))
events_indices = []
for i in xrange(len(events['left_lc_start'])):
t = (events['left_lc_start'][i], events['left_lc_end'][i], 'left_lane_change')
events_indices.append(t)
for i in xrange(len(events['right_lc_start'])):
t = (events['right_lc_start'][i], events['right_lc_end'][i], 'right_lane_change')
events_indices.append(t)
self.stats_file.close()
print events_indices
return events, events_indices