python opencv 全景图像拼接原理

前言 工作上的项目应用了surf计算特征。对surf进行初步的了解。
本篇参考链接：https://blog.csdn.net/u012483097/article/details/105369239/

1. 主要流程图
   
2. 灰度化图片

img_path1 = r"../left.jpg"
img_path2 = r"../right.jpg"
img1 = cv2.imread(img_path1)
img2 = cv2.imread(img_path2)
# 转化为灰度图
img1_gray = cv2.cvtColor(img1, cv2.COLOR_BGR2GRAY)
img2_gray = cv2.cvtColor(img2, cv2.COLOR_BGR2GRAY)

3. surf 计算

surf = cv2.xfeatures2d.SURF_create()
surf.create(1000)
# 计算img1， img2 图像的关键点和匹配点
kp1, des1 = surf.detectAndCompute(img1_gray, None)
kp2, des2 = surf.detectAndCompute(img2_gray, None)

4. 特征选择与匹配

'''
    BFMatcher.knnMatch() 暴力匹配  FlannBasedMatcher.knnMatch() flann匹配
    return : 这里的matches返回的是DMatch对象，DMatch有4个属性
    distance: 两个匹配点的距离，距离越短，匹配程度越高
    trainIdx: 训练描述符的索引
    queryIdx: 查询描述符的索引
    imgIdx: 训练图片的索引
    '''
    # ======================brute match============================#
    # 选择最优点
    goods = []
    img1points = []
    img2points = []
    if usemode == 'b':
        bf = cv2.BFMatcher()
        matches = bf.knnMatch(des1, des2, k=2)
    # ======================brute match end============================#

    # ======================knn match============================#
    elif usemode == 'k':
        kf = cv2.FlannBasedMatcher()
        matches = kf.knnMatch(des1, des2, 2)
    # ======================knn match end============================#
    # 特征匹配
    temp = cv2.drawMatches(img1, kp1, img2, kp2, goods, None)
    if len(goods) != 0:
        for i in goods:
            # queryIdx 表示查询点的位置 trainIdx表示匹配训练的点
            img1points.append(kp1[i.queryIdx].pt)
            img2points.append(kp2[i.trainIdx].pt)

5. 获取最优单应性变换矩阵

src_pts = np.float32([m for m in img1points])
dst_pts = np.float32([m for m in img2points])
# 在目标平面找需要匹配平面的最优单应性变换矩阵
H, _ = cv2.findHomography(dst_pts, src_pts, cv2.LMEDS)

6. 透视变换

img_warpPerspective = cv2.warpPerspective(img_right, H,
                                              (img_left.shape[1] + img_right.shape[1], min(img_left.shape[0],img_left.shape[0])))

7. 图片拼接

# 直接将左右拼接起来
w = img_left.shape[1]
img_warpPerspective[:, 0:w] = img_left

8. 结果

左图
右图拼接之后