computer-vision - 寻找拼接的单应性

Question

我正在处理以下任务：

我有 6 个鱼眼相机，想生成 360 度拼接图像。

在使用 findChessboardCorners、calibrateCamera 执行校准程序后，我获得了内在和外在矩阵。从6张带鱼眼效果的图像开始，通过fisheye.initUndistortRectifyMap函数，得到了6张平面图像。下面报告了上面的两个平面图像。

现在我应该进行拼接以获得 360 度图像。

我尝试使用 cv2.createStitcher 函数来做到这一点，但这并不总是有效，我也想访问单应矩阵来确定系统的静态矩阵。

所以我尝试计算单应矩阵，通过 SIFT 算法识别两幅图像之间的共同关键点，并保持最佳匹配的关键点。

然后我使用 warpPerspective 函数缝合了这两个图像。

我相信直到计算关键点的过程都是正确的，但我不明白为什么最终结果不好。

实际上，在尝试拼接第二张图像时，第二张图像在透视上完全变形/改变，失去了正确的图像。

这里有代码：

import cv2
import numpy as np


def cvshow(name, img):
    cv2.imshow(name, img)
    cv2.waitKey(0)
    cv2.destroyAllWindows()
    cv2.destroyAllWindows()


def sift_kp(image):
    gray_image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    sift = cv2.xfeatures2d.SIFT_create()
    sift = cv2.xfeatures2d.SIFT_create()
    kp, des = sift.detectAndCompute(image, None)
    kp_image = cv2.drawKeypoints(gray_image, kp, None)
    return kp_image, kp, des


def get_good_match(des1, des2):
    bf = cv2.BFMatcher()
    matches = bf.knnMatch(des1, des2, k=2)  # des1 is the template image, des2 is the matching image
    matches = sorted(matches, key=lambda x: x[0].distance / x[1].distance)
    good = []
    for m, n in matches:
        if m.distance < 0.55 * n.distance:
            good.append(m)
    return good


def drawMatches(imageA, imageB, kpsA, kpsB, matches, status):
    # Initialize the visualization picture, connect the A and B pictures left and right together
    (hA, wA) = imageA.shape[:2]
    (hB, wB) = imageB.shape[:2]
    vis = np.zeros((max(hA, hB), wA + wB, 3), dtype="uint8")
    vis[0:hA, 0:wA] = imageA
    vis[0:hB, wA:] = imageB

    # Joint traversal, draw matching pairs
    for ((trainIdx, queryIdx), s) in zip(matches, status):
        # When the point pair is matched successfully, draw it on the visualization
        if s == 1:
            # Draw matching pairs
            ptA = (int(kpsA[queryIdx][0]), int(kpsA[queryIdx][1]))
            ptB = (int(kpsB[trainIdx][0]) + wA, int(kpsB[trainIdx][1]))
            cv2.line(vis, ptA, ptB, (0, 255, 0), 1)

    # Return visualization results
    return vis


# Panorama stitching
def siftimg_rightlignment(img_right, img_left):
    _, kp1, des1 = sift_kp(img_right)
    _, kp2, des2 = sift_kp(img_left)
    goodMatch = get_good_match(des1, des2)
    # When the matching pairs of the filter items are greater than 4 pairs: calculate the perspective transformation matrix
    if len(goodMatch) > 4:
        # Get the point coordinates of the matching pair
        ptsA = np.float32([kp1[m.queryIdx].pt for m in goodMatch]).reshape(-1, 1, 2)
        ptsB = np.float32([kp2[m.trainIdx].pt for m in goodMatch]).reshape(-1, 1, 2)
        ransacReprojThreshold = 4
        H, status = cv2.findHomography(ptsA, ptsB, cv2.RANSAC, ransacReprojThreshold)

        print(H)
        #H = np.array([[-3.95002617e-01,-7.49813070e-02, 4.43642683e+02], [-4.06655962e-01,5.27365057e-01, 1.20636875e+02],[-1.60149798e-03, -3.69708507e-05, 1.00000000e+00]])

        # The function of this function is to first use RANSAC to select the best four sets of pairing points, and then calculate the H matrix. H is a 3*3 matrix

        # Change the angle of view to the right of the picture, result is the transformed picture
        result = cv2.warpPerspective(img_right, H, (img_right.shape[1] + img_left.shape[1], img_right.shape[0]))
        cvshow('result_medium', result)
        # Pass the picture left to the left end of the result picture
        result[0:img_left.shape[0], 0:img_left.shape[1]] = img_left
        return result


# Feature matching + panoramic stitching
import numpy as np
import cv2

# Read the stitched pictures (note the placement of the left and right pictures)
# Is to transform the graphics on the right
img_left = cv2.imread(r'\planar\0.png')
img_right = cv2.imread(r'\planar\5.png')

img_right = cv2.resize(img_right, None, fx=0.5, fy=0.3)
# Ensure that the two images are the same size
img_left = cv2.resize(img_left, (img_right.shape[1], img_right.shape[0]))

kpimg_right, kp1, des1 = sift_kp(img_right)
kpimg_left, kp2, des2 = sift_kp(img_left)

# Display the original image and the image after key point detection at the same time
cvshow('img_left', np.hstack((img_left, kpimg_left)))
cvshow('img_right', np.hstack((img_right, kpimg_right)))
goodMatch = get_good_match(des1, des2)
all_goodmatch_img = cv2.drawMatches(img_right, kp1, img_left, kp2, goodMatch, None, flags=2)

# goodmatch_img Set the first goodMatch[:10]
goodmatch_img = cv2.drawMatches(img_right, kp1, img_left, kp2, goodMatch[:10], None, flags=2)

cvshow('Keypoint Matches1', all_goodmatch_img)
cvshow('Keypoint Matches2', goodmatch_img)

# Stitch the picture into a panorama
result = siftimg_rightlignment(img_right, img_left)
cvshow('result', result)```

标签： computer-visionhomographyfisheyekeypointplanar-graph