python-opencv 直方图目标区域匹配
python-opencv 直方图目标区域匹配
1、python-opencv 直方图目标区域匹配
- 该方法可以匹配目标的大致区域，匹配过程中容易受颜色，光照，背景变化的影响
# -*- coding: utf-8 -*-
"""
Created on Tue Dec 27 09:32:02 2016
@author: http://blog.csdn.net/lql0716
"""
#首先确定目标区域的矩形框坐标，只需左上角的点和右下角的点的坐标，即mouseStart和mouseEnd，前一帧img1，当前帧img2
def histMatch(mouseStart, mouseEnd, img1, img2): #对相邻两帧的某个目标做直方图跟踪匹配， 仅仅是匹配目标的大致区域,传入的图片为彩色
a1 = mouseStart[1] #矩形区域左侧边 x坐标
a2 = mouseEnd[1] #矩形区域右侧边 x坐标
b1 = mouseStart[0] #矩形区域上侧边 y坐标
b2 = mouseEnd[0] #矩形区域下侧边 y坐标
height = mouseEnd[1] - mouseStart[1]
weight = mouseEnd[0] - mouseStart[0]
ratio = np.float(weight) / np.float(height) #weight与height的比例
# mask = np.zeros(img.shape[:2],np.uint8) #掩码操作，只对指定区域操作，这里初始值为 0，指定区域为 1
histsize = 256
#opencv方法读取-cv2.calcHist（速度最快）, 图像，通道[0]-灰度图，掩码-无，灰度级，像素范围
hist_full = cv2.calcHist([img1[a1:a2,b1:b2]],[0],None,[histsize],[0,256])
min = np.array([0.0, 0.0, 0.0, 0.0])
search_1 = 10 #搜索范围
search_2 = 0
if search_2 > 0:
#移动矩形区域左上角的点坐标，同时改变长和宽，进行搜索相似度最高的直方图
for i in range(-search_1, search_1+1):
for j in range(-search_1, search_1+1):
for k in range(-search_2, search_2+1):
if i == -search_1 and j == -search_1 and k == -search_2:
x1 = a1 + i
x2 = x1 + (height + k) #调整height的长度
y1 = b1 + j
y2 = y1 + (weight + int(round(k*ratio))) #调整weight的长度
else:
x1 = a1 + i
x2 = x1 + (height + k) #调整height的长度
y1 = b1 + j
y2 = y1 + (weight + int(round(k*ratio))) #调整weight的长度
hist_full2 = cv2.calcHist([img2[x1:x2, y1:y2]],[0],None,[histsize],[0,256])
comp = cv2.compareHist(hist_full, hist_full2, cv2.cv.CV_COMP_BHATTACHARYYA)
# print comp
if i == -search_1 and j == -search_1 and k == -search_2:
min[0]=comp
min[1]=i*1.0
min[2]=j*1.0
min[3]=k*1.0
# print 'a'
else:
if comp < min[0]:
min[0]=comp
min[1]=i*1.0
min[2]=j*1.0
min[3]=k*1.0
# print 'b'
else:
continue
#cv2.HISTCMP_BHATTACHARYYA 测试集测试，效果最好
#cv2.HISTCMP_CHISQR
#cv2.HISTCMP_CHISQR_ALT 效果第三
#cv2.HISTCMP_CORREL 效果很差
#cv2.HISTCMP_HELLINGER 效果第二
#cv2.HISTCMP_INTERSECT 效果很差
#cv2.HISTCMP_KL_DIV 效果很差
# cv2.cv.CV_COMP_BHATTACHARYYA 效果较好
# cv2.cv.CV_COMP_CHISQR 效果很差
# cv2.cv.CV_COMP_CORREL 效果比第一个稍差
# cv2.cv.CV_COMP_INTERSECT 效果很差
# print comp
print min
a = int(min[1]) #矩形左上角点的x坐标
b = int(min[2]) #矩形左上角点的y坐标
c = int(min[3]) #height的增量
new_a1 = a1 + a
new_a2 = new_a1 + (height + c)
new_b1 = b1 + b
new_b2 = b1 + (weight + int(round(c*ratio)))
new_h = height + c
new_w = weight + int(round(c*ratio))
# cv2.rectangle(img2,(new_b1,new_a1),(new_b2,new_a2),(0,0,255),1,8) #画矩形函数cv2.rectangle
new_mouse_start = np.array([new_b1, new_a1])
new_mouse_end = np.array([new_b2, new_a2])
return new_mouse_start, new_mouse_end
else:
#移动矩形区域左上角的点坐标，进行搜索相似度最高的直方图
for i in range(-search_1, search_1+1):
for j in range(-search_1, search_1+1):
if i == -search_1 and j == -search_1:
x1 = a1 + i
x2 = x1 + height
y1 = b1 + j
y2 = y1 + weight
else:
x1 = a1 + i
x2 = x1 + height
y1 = b1 + j
y2 = y1 + weight
hist_full2 = cv2.calcHist([img2[x1:x2, y1:y2]],[0],None,[histsize],[0,256])
comp = cv2.compareHist(hist_full, hist_full2, cv2.cv.CV_COMP_BHATTACHARYYA)
if i == -search_1 and j == -search_1:
min[0]=comp
min[1]=i*1.0
min[2]=j*1.0
# min[3]=k*1.0
# print 'a'
else:
if comp < min[0]:
min[0]=comp
min[1]=i*1.0
min[2]=j*1.0
# min[3]=k*1.0
# print 'b'
else:
continue
#cv2.HISTCMP_BHATTACHARYYA 测试集测试，效果最好
#cv2.HISTCMP_CHISQR
#cv2.HISTCMP_CHISQR_ALT 效果第三
#cv2.HISTCMP_CORREL 效果很差
#cv2.HISTCMP_HELLINGER 效果第二
#cv2.HISTCMP_INTERSECT 效果很差
#cv2.HISTCMP_KL_DIV 效果很差
# cv2.cv.CV_COMP_BHATTACHARYYA 效果较好
# cv2.cv.CV_COMP_CHISQR 效果很差
# cv2.cv.CV_COMP_CORREL 效果比第一个稍差
# cv2.cv.CV_COMP_INTERSECT 效果很差
# print comp
print min
a = int(min[1]) #矩形左上角点的x坐标
b = int(min[2]) #矩形左上角点的y坐标
new_a1 = a1 + a
new_a2 = new_a1 + height
new_b1 = b1 + b
new_b2 = new_b1 + weight
# cv2.rectangle(img2,(new_b1,new_a1),(new_b2,new_a2),(0,0,255),1,8) #画矩形函数cv2.rectangle
new_mouse_start = np.array([new_b1, new_a1])
new_mouse_end = np.array([new_b2, new_a2])
return new_mouse_start, new_mouse_end