我从汽车拍摄了视频。我的程序是测量前轮与道路白线之间的距离。该脚本对于左侧视频和右侧视频运行良好。
但是有时它会测量前轮和右侧白线之间的距离错误。
thresh = 150
distance_of_wood_plank = 80
pixel_of_wood_plank = 150
origin_width = 0
origin_height = 0
wheel_x = 0; wheel_y = 0 #xpoint and ypoint of wheel
df = pandas.DataFrame(columns=["Frame_No", "Distance", "TimeStrap"])
cap = cv2.VideoCapture(args.video)
frame_count = 0;
while(cap.isOpened()): #Reading input video by VideoCapture of Opencv
try:
frame_count += 1
ret, source = cap.read() # get frame from video
origin_height, origin_width, channels = source.shape
timestamps = [cap.get(cv2.CAP_PROP_POS_MSEC)]
milisecond = int(timestamps[0]) / 1000
current_time = str(datetime.timedelta(seconds = milisecond))
cv2.waitKey(1)
grayImage = cv2.cvtColor(source, cv2.COLOR_RGB2GRAY) # get gray image
crop_y = int(origin_height / 3 * 2) - 30
crop_img = grayImage[crop_y:crop_y + 100, 0:0 + origin_width] # get interest area
blur_image = cv2.blur(crop_img,(3,3))
ret, th_wheel = cv2.threshold(blur_image, 10, 255, cv2.THRESH_BINARY) #get only wheel
ret, th_line = cv2.threshold(blur_image, 150, 255, cv2.THRESH_BINARY) #get only white line
contours, hierarchy = cv2.findContours(th_wheel, cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)[-2:]
# get xpoint and ypoint of wheel
for cnt in contours:
x, y, w, h = cv2.boundingRect(cnt)
if (x < origin_width/ 4):
continue
elif (w < 10):
continue
elif (w > 80):
continue
elif (x > origin_width / 4 * 3):
continue
wheel_x = int(x)
wheel_y = int(y + h / 2 - 8)
pixel_count = 0 # count of pixel between wheel and white line
# get distance between wheel and white line
if (wheel_x > origin_width/2):
wheel_x -= 7
for i in range(wheel_x, 0, -1):
pixel_count += 1
suit_point = th_line[wheel_y,i]
if (suit_point == 255):
break
if (i == 1):
pixel_count = 0
pixel_count -= 4
cv2.line(source, (wheel_x - pixel_count, wheel_y + crop_y), (wheel_x, wheel_y + crop_y), (255, 0, 0), 2)
else :
wheel_x += 7
for i in range(wheel_x , origin_width):
pixel_count += 1
suit_point = th_line[wheel_y,i]
if (suit_point == 255):
break
if (i == origin_width - 1):
pixel_count = 0
pixel_count += 4
cv2.line(source, (wheel_x, wheel_y + crop_y), (wheel_x + pixel_count, wheel_y + crop_y), (255, 0, 0), 2)
distance_Cm = int(pixel_count * 80 / pixel_of_wood_plank)
str_distance = ""
if distance_Cm > 10:
str_distance = str(distance_Cm) + "Cm"
else:
str_distance = "No white line"
cv2.putText(source, str_distance, (50, 250), cv2.FONT_HERSHEY_SIMPLEX, 1, (255, 0, 0), 2, cv2.LINE_AA)
df = df.append({'Frame_No': frame_count,'Distance': str_distance ,'TimeStrap': current_time}, ignore_index = True)
df.to_csv("result.csv")
cv2.imshow("Distance_window", source)
cv2.waitKey(1)
except:
pass
这是视频的链接-https://drive.google.com/file/d/1IjJ-FA2LTGv8Cz-ReL7fFI7HPTiEhyxF/view?usp=sharing
答案 0 :(得分:1)
实际上,您在测量轮胎与白线之间的距离方面做得非常好。您需要考虑的是样品中有多少噪音。除非您停下卡车,下车,然后用胶带测量从轮胎到生产线的距离,否则您将永远无法真正知道轮胎的距离。您还需要考虑到(除非撞毁卡车)轮胎到白线的距离在每帧之间的变化不会超过几个像素。
最好的解决方案是卡尔曼滤波器,但这非常复杂。我使用了一个更简单的解决方案。为了找到行的位置,我对最后四个值取平均值以减少噪音。
import numpy as np, cv2
thresh = 150
distance_of_wood_plank = 80
pixel_of_wood_plank = 150
origin_width = 0
origin_height = 0
wheel_x = 0; wheel_y = 0 #xpoint and ypoint of wheel
cap = cv2.VideoCapture('/home/stephen/Desktop/20180301 1100 VW Right.mp4')
frame_count = 0;
vid_writer = cv2.VideoWriter('/home/stephen/Desktop/writer.avi', cv2.VideoWriter_fourcc('M','J','P','G'), 30, (480,360))
positions = []
import math
def distance(a,b): return math.sqrt((a[0]-b[0])**2 + (a[1]-b[1])**2)
while(cap.isOpened()): #Reading input video by VideoCapture of Opencv
frame_count += 1
ret, source = cap.read() # get frame from video
origin_height, origin_width, channels = source.shape
grayImage = cv2.cvtColor(source, cv2.COLOR_RGB2GRAY) # get gray image
crop_y = int(origin_height / 3 * 2) - 30
crop_img = grayImage[crop_y:crop_y + 100, 0:0 + origin_width] # get interest area
blur_image = cv2.blur(crop_img,(3,3))
ret, th_wheel = cv2.threshold(blur_image, 10, 255, cv2.THRESH_BINARY) #get only wheel
ret, th_line = cv2.threshold(blur_image, 150, 255, cv2.THRESH_BINARY) #get only white line
contours, hierarchy = cv2.findContours(th_wheel, cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)[-2:]
# get xpoint and ypoint of wheel
for cnt in contours:
x, y, w, h = cv2.boundingRect(cnt)
if (x < origin_width/ 4):
continue
elif (w < 10):
continue
elif (w > 80):
continue
elif (x > origin_width / 4 * 3):
continue
wheel_x = int(x)
wheel_y = int(y + h / 2 - 8)
pixel_count = 0 # count of pixel between wheel and white line
# get distance between wheel and white line
if (wheel_x > origin_width/2):
wheel_x -= 7
for i in range(wheel_x, 0, -1):
pixel_count += 1
suit_point = th_line[wheel_y,i]
if (suit_point == 255):
break
if (i == 1):
pixel_count = 0
pixel_count -= 4
else :
wheel_x += 7
for i in range(wheel_x , origin_width):
pixel_count += 1
suit_point = th_line[wheel_y,i]
if (suit_point == 255):
break
if (i == origin_width - 1):
pixel_count = 0
pixel_count += 4
a,b = (wheel_x - pixel_count, wheel_y + crop_y), (wheel_x, wheel_y + crop_y)
if distance(a,b)>10: positions.append((wheel_x + pixel_count, wheel_y + crop_y))
if len(positions)>10:
radius = 2
for position in positions[-10:]:
radius += 2
center = tuple(np.array(position, int))
color = 255,255,0
cv2.circle(source, center, radius, color, -1)
x,y = zip(*positions[-4:])
xa, ya = np.average(x), np.average(y)
center = int(xa), int(ya)
cv2.circle(source, center, 20, (0,0,255), 10)
cv2.imshow("Distance_window", source)
vid_writer.write(cv2.resize(source, (480,360)))
k = cv2.waitKey(1)
if k == 27: break
cv2.destroyAllWindows()