Skip to content
Snippets Groups Projects
board_detector.py 7.68 KiB
import cv2
import numpy as np

# global show_cv because I didn't want to have show_cv as an input to every function
show_cv = None
def init_show_cv(val):
    global show_cv
    show_cv = val

def find_longest_lines(img):
    gray_img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    
    # sobel gradients
    sobel_x = cv2.Sobel(gray_img, cv2.CV_64F, 1, 0, ksize=3)
    sobel_y = cv2.Sobel(gray_img, cv2.CV_64F, 0, 1, ksize=3)
    abs_sobel_x = np.absolute(sobel_x)
    abs_sobel_y = np.absolute(sobel_y)
    
    # threshold on abs values of sobel gradients and combine them
    _, threshold_x = cv2.threshold(abs_sobel_x, 17, 255, cv2.THRESH_BINARY)
    _, threshold_y = cv2.threshold(abs_sobel_y, 17, 255, cv2.THRESH_BINARY)
    combined_threshold = cv2.bitwise_or(threshold_x, threshold_y)
    combined_threshold = np.uint8(combined_threshold) # median blur needs this
    combined_threshold = cv2.medianBlur(combined_threshold, 5) # this gets rid of outliers so weird diagonal lines don't get made
    edges = combined_threshold
    # edges = cv2.Canny(gray_img, 30, 150, apertureSize=3) # didn't work as well

    if (show_cv):
        cv2.imshow('Sobel Filter', edges)
        cv2.waitKey(0)
        cv2.destroyAllWindows()

    lines = cv2.HoughLinesP(edges, 1, np.pi/180, 200, minLineLength=400, maxLineGap=15)

    vertical_lines = []
    horizontal_lines = []

    # separate horizontal and vertical lines
    if lines is not None:
        for line in lines:
            x1, y1, x2, y2 = line[0]
            if abs(x2 - x1) < abs(y2 - y1):  # vertical line
                vertical_lines.append(line)
            else:
                horizontal_lines.append(line)

    # filter lines too close to each other
    filtered_vertical = filter_lines(vertical_lines, 50)
    filtered_horizontal = filter_lines(horizontal_lines, 50)

    # sorted_vertical = sorted(filtered_vertical, key=lambda line: min(line[0][1], line[0][3]))
    # sorted_horizontal = sorted(filtered_horizontal, key=lambda line: min(line[0][0], line[0][2]))

    return filtered_vertical, filtered_horizontal

def filter_lines(lines, min_distance):
    filtered_lines = []

    # filter out lines too close to each other
    # (this assumes lines are around the same size and parallel)
    # (extremely simplified to improve computational speed because this is all we need)
    for line1 in lines:
        x1, y1, x2, y2 = line1[0]
        line1_x_avg = (x1 + x2) / 2
        line1_y_avg = (y1 + y2) / 2
        keep_line = True
        for line2 in filtered_lines:
            x3, y3, x4, y4 = line2[0]
            line2_x_avg = (x3 + x4) / 2
            line2_y_avg = (y3 + y4) / 2

            # calculate dist between average points of the 2 lines
            dist = np.sqrt((line1_x_avg - line2_x_avg)**2 + (line1_y_avg - line2_y_avg)**2)

            if dist < min_distance:
                keep_line = False
                break

        if keep_line:
            filtered_lines.append(line1)
            
    return filtered_lines


def detect_board(img):
    vertical_lines, horizontal_lines = find_longest_lines(img)
    print("# of Vertical:",len(vertical_lines))
    print("# of Horizontal:",len(horizontal_lines))

    height, width, _ = img.shape
    black_img = np.zeros((height, width), dtype=np.uint8)

    # create bitmasks for vert and horiz so we can get lines and intersections
    height, width, _ = img.shape
    vertical_mask = np.zeros((height, width), dtype=np.uint8)
    horizontal_mask = np.zeros((height, width), dtype=np.uint8)

    for line in vertical_lines:
        x1, y1, x2, y2 = line[0]
        cv2.line(vertical_mask, (x1, y1), (x2, y2), (255), 2)

    for line in horizontal_lines:
        x1, y1, x2, y2 = line[0]
        cv2.line(horizontal_mask, (x1, y1), (x2, y2), (255), 2)

    intersection = cv2.bitwise_and(vertical_mask, horizontal_mask)
    board_lines = cv2.bitwise_or(vertical_mask, horizontal_mask)
        
    contours, hierarchy = cv2.findContours(board_lines, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
    intersection_points, hierarchy = cv2.findContours(intersection, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)

    if (show_cv):
        board_lines_img = img.copy()
        cv2.drawContours(board_lines_img, contours, -1, (255, 255, 0), 2)
        cv2.drawContours(board_lines_img, intersection_points, -1, (0, 0, 255), 2)
        cv2.imshow('Lines of Board', board_lines_img)
        cv2.waitKey(0)
        cv2.destroyAllWindows()

    # find largest contour and get rid of it because it contains weird edges from lines
    max_area = 100000 # we're assuming board is going to be big (hopefully to speed up computation on raspberry pi)
    largest = -1
    # second_largest = -1
    # max_rect = None
    for i, contour in enumerate(contours):
        area = cv2.contourArea(contour)
        if area > max_area:
            max_area = area
            largest = i
    # "largest" is index of largest contour

    # get rid of contour containing the edges of the lines   
    contours = list(contours)
    contours.pop(largest)
    contours = tuple(contours)

    # thicken lines so that connections are made
    contour_mask = np.zeros((height, width), dtype=np.uint8)
    cv2.drawContours(contour_mask, contours, -1, (255), thickness=10)
    thick_contours, _ = cv2.findContours(contour_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

    # obtain largest contour of the thickened lines (the border) and approximate a 4 sided polygon onto it
    max_area = 100000
    largest = -1
    max_rect = None
    for i, contour in enumerate(thick_contours):
        area = cv2.contourArea(contour)
        if area > max_area:
            epsilon = 0.05 * cv2.arcLength(contour, True)
            rect = cv2.approxPolyDP(contour, epsilon, True) # uses Douglas-Peucker algorithm (probably overkill)
            if (len(rect) == 4):
                max_area = area
                largest = i
                max_rect = rect

    # perspective transform based on rectangle outline of board
    corners = max_rect.reshape(-1, 2) # turn rectangle into coordinate pairs
    tl = corners[1] # FIND A BETTER WAY TO DO THIS - sorting wasn't working for some reason
    tr = corners[0]
    bl = corners[2]
    br = corners[3]
    src = np.float32([list(tl), list(tr), list(bl), list(br)])
    dest = np.float32([[0,0], [width, 0], [0, height], [width, height]])
    M = cv2.getPerspectiveTransform(src, dest)
    Minv = cv2.getPerspectiveTransform(dest, src)
    warped_img = img.copy()
    warped_img = cv2.warpPerspective(np.uint8(warped_img), M, (width, height))

    M = cv2.getPerspectiveTransform(src, dest)
    Minv = cv2.getPerspectiveTransform(dest, src)
    warped_ip = img.copy()
    warped_ip = cv2.drawContours(warped_ip, intersection_points, -1, (0, 0, 255), 2)
    warped_ip = cv2.warpPerspective(np.uint8(warped_ip), M, (width, height))

    if (show_cv):
        contours_img = img.copy()
        # for i in range(63):
        #     cv2.drawContours(contours_img, [sorted_contours[i]], -1, (255-4*i, 4*i, 0), 2)
        cv2.drawContours(contours_img, thick_contours, -1, (0, 255, 0), 2)
        cv2.drawContours(contours_img, [thick_contours[largest]], -1, (0, 0, 255), 2)
        cv2.drawContours(contours_img, [max_rect], -1, (255, 0, 0), 2)
        for corner in corners:
            x,y = corner.ravel()
            cv2.circle(contours_img, (x, y), 5, (0, 255, 255), -1)
        # cv2.circle(contours_img, (int(min_x), int(min_y)), 5, (255, 0, 0), -1)
        # cv2.circle(contours_img, (int(max_x), int(max_y)), 5, (255, 0, 0), -1)
        cv2.imshow('Contours', contours_img)
        cv2.waitKey(0)
        cv2.destroyAllWindows()

        cv2.imshow('Warped', warped_img)
        cv2.waitKey(0)
        cv2.destroyAllWindows()

        # cv2.imshow('Warped', warped_ip)
        # cv2.waitKey(0)
        # cv2.destroyAllWindows()