准备

1、一张棋盘图

可以直接从opencv官方github下载,这是一个拥有10*7个格子的棋盘,共有9*6个角点,每个格子24mm,本文所使用的就是这一个棋盘。你需要将它打印在A4纸上用于后续使用。(也可以根据官方教程自行设置棋盘大小OpenCV: Create calibration pattern)

opencv/pattern.png at 4.x · opencv/opencv · GitHub

2、一个双目摄像头

随便在tb买的一个不知名摄像头,附赠了一个.exe的测试工具用于简单使用摄像头效果如下

使用opencv简单测试一下,我用的笔记本,接上usb摄像头就是从1开始了,这个双目摄像头虽然有两个输入index=1和index=2但是其实只需要获取index=1的那个视频流就可以得到双目效果。

import cv2cap = cv2.VideoCapture(1)cap.set(cv2.CAP_PROP_FRAME_WIDTH,1280)cap.set(cv2.CAP_PROP_FRAME_HEIGHT,480)while(1):_, frame = cap.read()assert _, "摄像头获取失败"cv2.imshow('img', frame)c = cv2.waitKey(1)if c == 27:cap.release()break

开启前必须将分辨率设置为正确的宽度,我的相机是1280,如果设置宽度不正确会导致无法正确得到双目图像

可以通过下面代码获取相机分辨率,主要是获得width,双目图的width应该为两个相机的width之和

import cv2cap0 = cv2.VideoCapture(1)cap1 = cv2.VideoCapture(2)res0 = [cap0.get(cv2.CAP_PROP_FRAME_WIDTH),cap0.get(cv2.CAP_PROP_FRAME_HEIGHT)]res1 = [cap1.get(cv2.CAP_PROP_FRAME_WIDTH),cap1.get(cv2.CAP_PROP_FRAME_HEIGHT)] print(res0)print(res1)cap0.release()cap1.release()

分辨率正确的双目图(1280*480)

分辨率错误的双目图(2560*480)

开始操作

先给棋盘拍照

import cv2,oscap = cv2.VideoCapture(1)cap.set(cv2.CAP_PROP_FRAME_WIDTH,1280)cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 480)# 生成目录path = './calibration/'path_l = './calibration/left/'path_r = './calibration/right/'os.mkdir(path) if not os.path.exists(path) else Noneos.mkdir(path_l) if not os.path.exists(path_l) else Noneos.mkdir(path_r) if not os.path.exists(path_r) else Nonecount = 0while cap.isOpened():ret, frame = cap.read()cv2.imshow('img',frame)k = cv2.waitKey(1)# 按下ESC退出if k == 27:break# 按下空格键暂停if k == 32:cv2.imshow('img',frame)# 再次按下空格保存if cv2.waitKey() == 32:cv2.imwrite(path + "{}.jpg".format(count), frame)# 保存全图cv2.imwrite(path_l + "{}.jpg".format(count), frame[:,0:640])# 保存左图cv2.imwrite(path_r + "{}.jpg".format(count), frame[:,640:])# 保存右图count += 1cv2.destroyAllWindows()cap.release()

按照下图至少拍摄12对左右图像,以获得最佳效果

来源:Stereo Calibration for the Dual Camera Mezzanine – Blog – FPGA – element14 Community

测试一下棋盘角点绘制

import cv2img = cv2.imread('./calib/left/0.jpg')img1 = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)ret, corner = cv2.findChessboardCorners(img1, (9,6))ret, corner = cv2.find4QuadCornerSubpix(img1, corner, (7,7))cv2.drawChessboardCorners(img, (9,6), corner, ret)cv2.imshow('corner', img)cv2.waitKey(0)

接下来就获取矫正所需要的参数

import cv2, globimport numpy as np'''获得标定所需参数'''# 定义棋盘格的大小chessboard_size = (9, 6)# 定义图像分辨率,根据自己相机的分辨率修改imgsz = (640, 480)# 定义棋盘格中每个格子的物理大小,自己用尺子量,单位为毫米(mm)square_size = 24# 定义棋盘格模板的点的坐标objp = np.zeros((chessboard_size[0]*chessboard_size[1], 3), np.float32) #生成每个角点三维坐标,共有chessboard_size[0]*chessboard_size[1]个坐标,z轴置0不影响objp[:, :2] = np.mgrid[0:chessboard_size[0], 0:chessboard_size[1]].T.reshape(-1, 2) * square_size #计算得到每个角点的x,y# 读取所有棋盘格图像并提取角点imgpoints_left, imgpoints_right = [], []# 存储图像中的角点objpoints = []# 存储模板中的角点images = glob.glob('./calibration/right/*.jpg')# 所有棋盘格图像所在的目录for fname in images:img = cv2.imread(fname)gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)ret, corners = cv2.findChessboardCorners(gray, chessboard_size, None) #计算cornerret, corners = cv2.find4QuadCornerSubpix(gray, corners, (7,7)) #提高角点检测的准确性和稳定性if ret == True:imgpoints_right.append(corners)objpoints.append(objp)images = glob.glob('./calibration/left/*.jpg')# 所有棋盘格图像所在的目录for fname in images:img = cv2.imread(fname)gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)ret, corners = cv2.findChessboardCorners(gray, chessboard_size, None) #计算cornerret, corners = cv2.find4QuadCornerSubpix(gray, corners, (7,7)) #提高角点检测的准确性和稳定性if ret == True:imgpoints_left.append(corners)'''开始标定,获得参数'''# 标定相机,获得内参和畸变参数ret, mtx_r, dist_r, rvecs_r, tvecs_r = cv2.calibrateCamera(objpoints, imgpoints_right, gray.shape[::-1], None, None)ret, mtx_l, dist_l, rvecs_l, tvecs_l = cv2.calibrateCamera(objpoints, imgpoints_left, gray.shape[::-1], None, None)# 指定迭代次数最大30或者误差小于0.001term = (cv2.TERM_CRITERIA_EPS + cv2.TERM_CRITERIA_MAX_ITER, 30, 0.001)# 进行双目相机标定,主要是获得R,T两个矩阵rotation_matrix, translation_matrix = cv2.stereoCalibrate(objpoints, imgpoints_left, imgpoints_right,mtx_l, dist_l,mtx_r, dist_r,imgsz, flags=cv2.CALIB_FIX_INTRINSIC, criteria=term)[5:7]# 获得矫正矩阵和投影矩阵,用于后续进行图像校正rect_left, rect_right, \proj_left, proj_right, \dispartity, \ROI_left, ROI_right = cv2.stereoRectify(mtx_l, dist_l,mtx_r, dist_r,imgsz, rotation_matrix, translation_matrix,flags=cv2.CALIB_ZERO_DISPARITY, alpha=-1)'''打印结果'''print('mtx_l = np.array({})'.format(np.array2string(mtx_l, separator=', ', formatter={'int': lambda x: f'{x: 3d}'},prefix='[', suffix=']')))print('mtx_r = np.array({})'.format(np.array2string(mtx_r, separator=', ', formatter={'int': lambda x: f'{x: 3d}'},prefix='[', suffix=']')))print('dist_l = np.array({})'.format(np.array2string(dist_l, separator=', ', formatter={'int': lambda x: f'{x: 3d}'},prefix='[', suffix=']')))print('dist_r = np.array({})'.format(np.array2string(dist_r, separator=', ', formatter={'int': lambda x: f'{x: 3d}'},prefix='[', suffix=']')))print('R = np.array({})'.format(np.array2string(rotation_matrix, separator=', ', formatter={'int': lambda x: f'{x: 3d}'},prefix='[', suffix=']')))print('T = np.array({})'.format(np.array2string(translation_matrix, separator=', ', formatter={'int': lambda x: f'{x: 3d}'},prefix='[', suffix=']')))print('rect_left = np.array({})'.format(np.array2string(rect_left, separator=', ', formatter={'int': lambda x: f'{x: 3d}'},prefix='[', suffix=']')))print('rect_right = np.array({})'.format(np.array2string(rect_right, separator=', ', formatter={'int': lambda x: f'{x: 3d}'},prefix='[', suffix=']')))print('proj_left = np.array({})'.format(np.array2string(proj_left, separator=', ', formatter={'int': lambda x: f'{x: 3d}'},prefix='[', suffix=']')))print('proj_right = np.array({})'.format(np.array2string(proj_right, separator=', ', formatter={'int': lambda x: f'{x: 3d}'},prefix='[', suffix=']')))print('dispartity = np.array({})'.format(np.array2string(dispartity, separator=', ', formatter={'int': lambda x: f'{x: 3d}'},prefix='[', suffix=']')))# print('mtx_l = np.array({})'.format(mtx_l))# print('mtx_r = np.array({})'.format(mtx_r))# print('dist_l = np.array({})'.format(dist_l))# print('dist_r = np.array({})'.format(dist_r))# print('R = np.array({})'.format(rotation_matrix))# print('T = np.array({})'.format(translation_matrix))# print('rect_left = np.array({})'.format(rect_left))# print('rect_right = np.array({})'.format(rect_right))# print('proj_left = np.array({})'.format(proj_left))# print('proj_right = np.array({})'.format(proj_right))# print('dispartity = np.array({})'.format(dispartity))print('ROI_left = np.array({})'.format(ROI_left))print('ROI_right = np.array({})'.format(ROI_right))

得到下面参数

可以直接复制用于图像矫正

测试

import cv2, glob, osimport numpy as npdef get_corners(imgs, corners):for img in imgs:# 9x12棋盘有8x11个角点ret, c = cv2.findChessboardCorners(img, (9,6))assert(ret)ret, c = cv2.find4QuadCornerSubpix(img, c, (7,7))assert(ret)corners.append(c)mtx_l = np.array([[479.61836296, 0., 339.91341613], [0., 478.44413757, 240.61069496], [0., 0., 1.,]])mtx_r = np.array([[483.4989366,0., 306.98497259], [0., 482.17064224, 228.91672333], [0., 0., 1.]])dist_l = np.array([[ 0.07539615, -0.51291496,0.00405133, -0.00084347,0.7514282 ]])dist_r = np.array([[-1.30834008e-01,8.25592192e-01,9.83305297e-04, -7.40611932e-06,-1.67568022e+00]])R = np.array([[ 9.99947786e-01, -1.06501500e-03,1.01632001e-02], [ 8.52847758e-04,9.99782093e-01,2.08575744e-02], [-1.01831991e-02, -2.08478176e-02,9.99730799e-01]])T = np.array([[-62.0710667 ], [0.27233791], [0.49530174]])rect_left = np.array([[ 0.99998384, -0.005285,0.00209416], [ 0.00526285,0.99993159,0.01044553], [-0.00214922, -0.01043434,0.99994325]])rect_right = np.array([[ 0.99995854, -0.00438734, -0.00797926], [ 0.00430379,0.99993606, -0.01045726], [ 0.00802463,0.01042249,0.99991348]])proj_left = np.array([[480.3073899,0., 322.84606934, 0.,], [0., 480.3073899,235.60386848, 0.,], [0., 0., 1., 0.,]])proj_right = np.array([[ 4.80307390e+02,0.00000000e+00,3.22846069e+02, -2.98144281e+04], [ 0.00000000e+00,4.80307390e+02,2.35603868e+02,0.00000000e+00], [ 0.00000000e+00,0.00000000e+00,1.00000000e+00,0.00000000e+00]])dispartity = np.array([[ 1.00000000e+00,0.00000000e+00,0.00000000e+00, -3.22846069e+02], [ 0.00000000e+00,1.00000000e+00,0.00000000e+00, -2.35603868e+02], [ 0.00000000e+00,0.00000000e+00,0.00000000e+00,4.80307390e+02], [ 0.00000000e+00,0.00000000e+00,1.61098978e-02, -0.00000000e+00]])ROI_left = np.array((5, 10, 612, 456))ROI_right = np.array((14, 5, 626, 475))img_left = []img_right = []corners_left = []corners_right = []img_file = glob.glob('./calibration/*.jpg')imgsize = (640, 480)for img in img_file:frame = cv2.imread(img)frame = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)l = frame[:,0:640]r = frame[:,640:]img_left.append(l)img_right.append(r)print("获取角点", "left")get_corners(img_left, corners_left)print("获取角点", "right")get_corners(img_right, corners_right)for i in range(len(img_left)):l = img_left[i]r = img_right[i]# 计算双目校正的矩阵R1, R2, P1, P2, Q, validPixROI1, validPixROI2 = cv2.stereoRectify(mtx_l, dist_l, mtx_r, dist_r, imgsize, R, T)# 计算校正后的映射关系maplx , maply = cv2.initUndistortRectifyMap(mtx_l, dist_l, R1, P1, imgsize, cv2.CV_16SC2)maprx , mapry = cv2.initUndistortRectifyMap(mtx_r, dist_r, R2, P2, imgsize, cv2.CV_16SC2)# 映射新图像lr = cv2.remap(l, maplx, maply, cv2.INTER_LINEAR)rr = cv2.remap(r, maprx, mapry, cv2.INTER_LINEAR)all = np.hstack((lr,rr))# 变换之后和变换之前的角点坐标不一致,所以线不是正好经过角点,只是粗略估计,但偶尔能碰到离角点比较近的线,观察会比较明显cv2.line(all, (-1, int(corners_left[i][0][0][1])), (all.shape[1], int(corners_left[i][0][0][1])), (255), 1)# 可以看出左右图像y坐标对齐还是比较完美的,可以尝试着打印双目校正前的图片,很明显,左右y坐标是不对齐的cv2.imshow('a', all)c = cv2.waitKey()cv2.destroyAllWindows()if c == 27:breakprint("end")

此段代码借鉴http://t.csdn.cn/uxwLA

参考链接:

Depther project – part 2: calibrate dual camera, parameters rectification – edgenoon.ai

http://t.csdn.cn/uxwLA

OpenCV: Create calibration pattern

opencv/pattern.png at 4.x · opencv/opencv · GitHub

Stereo Calibration for the Dual Camera Mezzanine – Blog – FPGA – element14 Community