如何解决从多张照片中的棋盘格网格估计相机姿势时确保共享原点
我有两张静止场景的照片,其中包括棋盘格图案和瓶子。我的目标是估计相对于共享原点的相机姿势,以便我可以执行进一步的处理。我最初的计划是使用棋盘格来估计相机位置。我对这个过程感到担忧,因为棋盘格的原点在两个图像中可能不同。即,由于棋盘图案的对称性,被选为棋盘图案原点的点是否可能因照片而异。如果是这样,无论如何要避免这种情况,同时仍然使用棋盘格网格进行姿势估计。
我正在使用下面的代码推导出我的相机的姿势
################### CORNER FINDING CRITERIA ####################
criteria = (cv.TERM_CRITERIA_EPS + cv.TERM_CRITERIA_MAX_ITER,30,0.001)
################### GENERATE THE OBJECT POINTS #################
squareSZ = 23.876 #Square edge length in mm
dim1 = 9
dim2 = 6
objp = np.zeros((dim2*dim1,3),np.float32) #Generates object points
objp[:,:2] = np.mgrid[0:dim1,0:dim2].T.reshape(-1,2)*squareSZ #Generates a 3d grid from object points
################### LOAD CAMERA EXTRINSICS #####################
camMat = np.array(pickle.load(open('cameraMatrix_GoPro.sav','rb'))) #Loads the camera matrix from the calibration
distCoeffs = np.array(pickle.load(open('distCoeff_GoPro.sav','rb'))) #Loads the distortion coefficients from calibration
################### LOAD CAMERA IMAGES #########################
file1 = 'WaterBottleTooSide/GOPR0250.JPG'
file2 = 'WaterBottleTooSide/GOPR0251.JPG'
# Loads Images
img1 = cv.imread(file1)
img2 = cv.imread(file2)
# Converts to grayscale
img1Gray = cv.cvtColor(img1,cv.COLOR_BGR2GRAY)
img2Gray = cv.cvtColor(img2,cv.COLOR_BGR2GRAY)
################### CALCULATE CAMERA EXTRINSICS #################
ret1,camCorners1 = cv.findChessboardCorners(img1Gray,(dim1,dim2),None) # Find the chess board corners
ret2,camCorners2 = cv.findChessboardCorners(img2Gray,None)
if ret1 == True:
#Iterates to find the sub-pixel accurate location of corners or radial saddle points
camPoints1 = np.squeeze(cv.cornerSubPix(img1Gray,camCorners1,(11,11),(-1,-1),criteria))
worldPoints=np.squeeze(np.array(objp)) #Points from my world frame
retval,rvecs1,tvecs1 = cv.solvePnP(worldPoints,camPoints1,camMat,distCoeffs) #Calculates location using camera and world view points
rotMat1 = cv.Rodrigues(rvecs1)[0] #Calculates rotation matrix from the rvec
R_t1 = np.hstack((rotMat1,tvecs1)) #Calculates the 3x4 matrix [R|t]
P1 = np.matmul(camMat,R_t1) #Calculates the projection matrix
P1inv = np.linalg.pinv(P1) #Calculate the pseudo-inverse of the projection matrix 4x3 matrix
C1 = np.matmul(-np.transpose(rotMat1),tvecs1) #Location of camera center for first image
C1 = np.vstack((C1,1))#Concatenates 1 to create homogenous point
if ret2 == True:
#Iterates to find the sub-pixel accurate location of corners or radial saddle points
camPoints2 = np.squeeze(cv.cornerSubPix(img2Gray,camCorners2,rvecs2,tvecs2 = cv.solvePnP(worldPoints,camPoints2,distCoeffs) #Calculates location using camera and world view points
rotMat2 = cv.Rodrigues(rvecs2)[0] #Calculates rotation matrix from the rvec
R_t2 = np.hstack((rotMat2,tvecs2)) #Calculates the 3x4 matrix [R|t]
P2 = np.matmul(camMat,R_t2) #Calculates the projection matrix
P2inv = np.linalg.pinv(P2) #Calculate the pseudo-inverse of the projection matrix 4x3 matrix
C2 = np.matmul(-np.transpose(rotMat2),tvecs2) #Location of camera center for first image
C2 = np.vstack((C2,1))#Concatenates 1 to create homogenous point
版权声明:本文内容由互联网用户自发贡献,该文观点与技术仅代表作者本人。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如发现本站有涉嫌侵权/违法违规的内容, 请发送邮件至 dio@foxmail.com 举报,一经查实,本站将立刻删除。