-
Notifications
You must be signed in to change notification settings - Fork 0
/
image_manipulator.py
224 lines (180 loc) · 6.86 KB
/
image_manipulator.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
from math import tan, atan, sin, pi, radians, sqrt
import time
import json
import types
import cv2
import numpy as np
def getTargetHeight(targetWidth, cameraInclination, fovy, yOnScreen):
alpha1 = pi/2 - cameraInclination
alpha2 = pi/2 - atan(tan(fovy/2) * (1-2*yOnScreen))
#print(alpha1, alpha2, pi-alpha1-alpha2, width, yOnScreen)
return targetWidth * (sin(alpha2) / sin(pi-alpha1-alpha2) * yOnScreen)
def warpImage(image, corners, targetWidth, targetHeight):
corners = np.array(corners, dtype=np.float32)
#print(corners)
#corners = order_points(corners)
target = [(0, 0), (targetWidth-1, 0), (targetWidth-1, targetHeight-1), (0, targetHeight-1)]
target = np.array(target, dtype=np.float32)
mat = cv2.getPerspectiveTransform(corners, target)
out = cv2.warpPerspective(image, mat, (targetWidth, targetHeight))
return out
def getStreetRect(img, cameraInclination, fovy, upperRectLineHeight, targetWidth):
height, width, _ = np.shape(img)
screenRatio = width / height
tanLineAngle = (tan(cameraInclination) / tan(fovy/2) + 1) / screenRatio
def f(x):
return tanLineAngle * x
def fInverse(y):
return y / tanLineAngle
if f(upperRectLineHeight) > 1/screenRatio:
x, y = fInverse(1/screenRatio), 1/screenRatio
else:
x, y = upperRectLineHeight, f(upperRectLineHeight)
corners = [[x*width, height - y*width], [(1-x)*width, height - y*width],
[width, height], [0, height]]
targetHeight = int(getTargetHeight(targetWidth, cameraInclination, fovy, y*screenRatio))
return corners, warpImage(img, corners, targetWidth, targetHeight)
integerInfinity = 1<<31 - 1 # max value for 32bit signed ints (needed in numpy)
def getRoadRadiuses(count, multiplier):
"""returns 2*count radiuses sorted from bigger to smaller"""
positives = []
negatives = []
for i in range(count):
radius = int(multiplier * count / (i+1))
if radius > integerInfinity:
radius = integerInfinity
elif radius < 2:
radius = 2
positives.append(radius)
negatives.append(-radius)
return positives + negatives[::-1]
def enumerateRadiuses(radiuses):
"""yields all correspodingIndex,radiuses and also None,infinity"""
yield None, integerInfinity
for radius in enumerate(radiuses):
yield radius
def circularShift(src, r, roadDistance):
direction = 1 if r < 0 else -1
r = abs(r)
if r <= roadDistance:
return None
height, width, channels = np.shape(src)
cumulativeHeight = height + roadDistance
if r <= width:
if r <= cumulativeHeight:
newHeight = r
else:
newHeight = cumulativeHeight
else:
if r <= cumulativeHeight:
newHeight = int(sqrt(2*r*width - width*width))
else:
newHeight = min(cumulativeHeight, int(sqrt(2*r*width - width*width)))
newHeight -= roadDistance
shifted = np.full((newHeight, width, channels), -1, dtype=np.int16)
for h in range(1, newHeight+1):
shift = int(r - sqrt(r*r - (h + roadDistance)**2))
if direction == 1:
shifted[newHeight-h, shift:width] = src[newHeight-h, 0:width-shift]
else:
shifted[newHeight-h, 0:width-shift] = src[newHeight-h, shift:width]
return shifted
def columnAverage(src):
_, width, _ = np.shape(src)
average = np.zeros((width,), dtype=np.uint8)
for w in range(width):
average[w] = np.average(src[:, w, :], weights=((src[:, w, :] >= 0) + [1e-10]))
return average
def pruneColumnAverage(src, average, count):
"""in lines where only <=count pixels are valid, the values can't be considered ok"""
_, width, _ = np.shape(src)
for w in range(2*width//3, width):
countValid = np.sum(src[:, w, :] >= 0)
if countValid <= count:
average[w] = average[w-1]
for w in range(width//3, -1, -1):
countValid = np.sum(src[:, w, :] >= 0)
if countValid <= count:
average[w] = average[w+1]
return average
def maxDifference(arr):
diff = -1
for i in range(8, len(arr)-9):
diff = max(diff, arr[i]-arr[i+1] if arr[i] > arr[i+1] else arr[i+1]-arr[i])
return diff
def processImage(src, p):
start = time.time()
_, rect = getStreetRect(src, p.cameraInclination, p.fovy,
p.upperRectLineHeight, p.profileWidth)
#cv2.imshow("rect", rect)
#cv2.imshow("frame", frame)
#print(corn)
#prev = corn[-1]
#for c in corn:
# cv2.circle(img, (int(c[0]), int(c[1])), radius=5, color=(255,255,255,255), thickness=7)
# cv2.line(img, (int(prev[0]), int(prev[1])), (int(c[0]), int(c[1])),
# color=(255,255,255,255), thickness=2)
# prev = c
bestDiff = -1
bestRadius = None
bestShift = None
bestAverage = None
for _, radius in enumerateRadiuses(processImage.radiuses):
shifted = circularShift(rect, radius, 49)
if shifted is None:
continue
average = columnAverage(shifted)
average = pruneColumnAverage(shifted, average, 10)
diff = maxDifference(average)
if (diff > bestDiff) or (diff == bestDiff and abs(radius) > abs(bestRadius)):
bestDiff = diff
bestRadius = radius
bestShift = shifted
bestAverage = average
end = time.time()
print(end - start, "s")
cv2.imshow("bestShift", bestShift.astype(np.uint8))
cv2.imshow("bestAverage", arrToImg(bestAverage))
print(bestRadius)
def getParams():
res = types.SimpleNamespace()
data = json.load(open("./params.json"))
res.width = data["width"]
res.height = data["height"]
res.cameraInclination = radians(data["cameraInclination"])
res.cameraHeight = data["cameraHeight"]
res.upperRectLineHeight = data["upperRectLineHeight"]
res.profileWidth = data["profileWidth"]
res.videoPath = data["videoPath"]
res.datasetPath = data["datasetPath"]
res.fovx = radians(data["fovx"])
res.fovy = 2 * atan(tan(res.fovx/2) / res.width * res.height)
return res
def arrToImg(a, highlight=None):
width = np.shape(a)[0]
res = np.empty((5, width, 3), dtype=np.uint8)
for i in range(width):
res[:, i, :] = a[i]
if highlight is not None:
res[:, highlight, 0] = 0
return res
def main():
p = getParams()
processImage.radiuses = getRoadRadiuses(40, 5 * p.profileWidth)
cap = cv2.VideoCapture(p.videoPath)
while not cap.isOpened():
cap = cv2.VideoCapture(p.videoPath)
cv2.waitKey(1000)
i = 0
while True:
flag, frame = cap.read()
if flag:
if i%30 == 0:
processImage(frame, p)
i += 1
if cv2.waitKey(1) == 27:
break
if cap.get(cv2.CAP_PROP_POS_FRAMES) == cap.get(cv2.CAP_PROP_FRAME_COUNT):
break
if __name__ == "__main__":
main()