superimpose.py

import numpy as np


def get_h_matrices(poly_curve, x, y, orientation=0):
    """
    get H-matrix for detected orientation
    :param poly_curve: the approximate poly curve yielded by approxPolyDP
    :param x: no. of rows in the image frame
    :param y: no. of columns in the image frame
    :param orientation: orientation of the tag represented using integers 0 to 3
    :return: the homogeneous or inverse-homogeneous transform matrix
    """
    orientations = {'bottom_right': 0, 'bottom_left': 1, 'top_right': 2, 'top_left': 3}
    x_width = np.zeros(4, dtype=int)
    y_width = np.zeros(4, dtype=int)
    x_center = np.array([0, x, x, 0])
    y_center = np.array([0, 0, y, y])

    # Define width to perform homogeneous transforms
    if orientation == orientations['bottom_right']:
        x_width[0], y_width[0] = poly_curve[0][0][0], poly_curve[0][0][1]
        x_width[1], y_width[1] = poly_curve[1][0][0], poly_curve[1][0][1]
        x_width[2], y_width[2] = poly_curve[2][0][0], poly_curve[2][0][1]
        x_width[3], y_width[3] = poly_curve[3][0][0], poly_curve[3][0][1]
    elif orientation == orientations['bottom_left']:
        x_width[0], y_width[0] = poly_curve[1][0][0], poly_curve[1][0][1]
        x_width[1], y_width[1] = poly_curve[2][0][0], poly_curve[2][0][1]
        x_width[2], y_width[2] = poly_curve[3][0][0], poly_curve[3][0][1]
        x_width[3], y_width[3] = poly_curve[0][0][0], poly_curve[0][0][1]
    elif orientation == orientations['top_right']:
        x_width[0], y_width[0] = poly_curve[2][0][0], poly_curve[2][0][1]
        x_width[1], y_width[1] = poly_curve[3][0][0], poly_curve[3][0][1]
        x_width[2], y_width[2] = poly_curve[0][0][0], poly_curve[0][0][1]
        x_width[3], y_width[3] = poly_curve[1][0][0], poly_curve[1][0][1]
    elif orientation == orientations['top_left']:
        x_width[0], y_width[0] = poly_curve[3][0][0], poly_curve[3][0][1]
        x_width[1], y_width[1] = poly_curve[0][0][0], poly_curve[0][0][1]
        x_width[2], y_width[2] = poly_curve[1][0][0], poly_curve[1][0][1]
        x_width[3], y_width[3] = poly_curve[2][0][0], poly_curve[2][0][1]
    else:
        print('Incorrect Orientation!!')
        quit()

    # Evaluate the A matrix
    a_mat = [[x_width[0], y_width[0], 1, 0, 0, 0, -x_center[0] * x_width[0], -x_center[0] * y_width[0], -x_center[0]],
             [0, 0, 0, x_width[0], y_width[0], 1, -y_center[0] * x_width[0], -y_center[0] * y_width[0], -y_center[0]],
             [x_width[1], y_width[1], 1, 0, 0, 0, -x_center[1] * x_width[1], -x_center[1] * y_width[1], -x_center[1]],
             [0, 0, 0, x_width[1], y_width[1], 1, -y_center[1] * x_width[1], -y_center[1] * y_width[1], -y_center[1]],
             [x_width[2], y_width[2], 1, 0, 0, 0, -x_center[2] * x_width[2], -x_center[2] * y_width[2], -x_center[2]],
             [0, 0, 0, x_width[2], y_width[2], 1, -y_center[2] * x_width[2], -y_center[2] * y_width[2], -y_center[2]],
             [x_width[3], y_width[3], 1, 0, 0, 0, -x_center[3] * x_width[3], -x_center[3] * y_width[3], -x_center[3]],
             [0, 0, 0, x_width[3], y_width[3], 1, -y_center[3] * x_width[3], -y_center[3] * y_width[3], -y_center[3]]]
    # Get inverse homogeneous transform using svd
    _, _, v_h = np.linalg.svd(a_mat, full_matrices=True)
    h_mat = np.array(v_h[8, :] / v_h[8, 8]).reshape((-1, 3))
    inv_h = np.linalg.inv(h_mat)
    # Return inverse homogeneous transform
    return h_mat, inv_h


def get_warp_perspective(transpose_image, h_matrix, dimension):
    """
    get image warped onto the video frame
    :param transpose_image: transpose of the image to be warped
    :param h_matrix: homography matrix: use inverse homography matrix to warp given image onto the video
    :param dimension: a tuple of dimension (rows, columns) of image to be warped in; reverse the tuple to warp a given
    image onto the video
    :return: warped image; NOTE - take transpose of this image to see output
    """
    warped_image = np.zeros((dimension[0], dimension[1], 3))
    for index1 in range(0, transpose_image.shape[0]):
        for index2 in range(0, transpose_image.shape[1]):
            new_vec = np.dot(h_matrix, [index1, index2, 1])
            new_row, new_col, _ = (new_vec / new_vec[2] + 0.4).astype(int)
            if 5 < new_row < (dimension[0] - 5):
                if 5 < new_col < (dimension[1] - 5):
                    warped_image[new_row, new_col] = transpose_image[index1, index2]
                    warped_image[new_row - 1, new_col - 1] = transpose_image[index1, index2]
                    warped_image[new_row - 2, new_col - 2] = transpose_image[index1, index2]
                    warped_image[new_row - 3, new_col - 3] = transpose_image[index1, index2]
                    warped_image[new_row + 1, new_col + 1] = transpose_image[index1, index2]
                    warped_image[new_row + 2, new_col + 2] = transpose_image[index1, index2]
                    warped_image[new_row + 3, new_col + 3] = transpose_image[index1, index2]

    return np.array(warped_image, dtype=np.uint8)