python+opencv车道线检测（简易实现）

python+opencv车道线检测（简易实现）

技术栈：python+opencv

实现思路：

1. canny边缘检测获取图中的边缘信息；
2. 霍夫变换寻找图中直线；
3. 绘制梯形感兴趣区域获得车前范围；
4. 得到并绘制车道线；

效果展示：

代码实现：

import cv2
import numpy as np


def canny():
    gray = cv2.cvtColor(lane_image, cv2.COLOR_RGB2GRAY)
    #高斯滤波
    blur = cv2.GaussianBlur(gray, (5, 5), 0)
	#边缘检测
    canny_img = cv2.Canny(blur, 50, 150)
    return canny_img


def region_of_interest(r_image):
    h = r_image.shape[0]
    w = r_image.shape[1]
	# 这个区域不稳定，需要根据图片更换
    poly = np.array([
        [(100, h), (500, h), (290, 180), (250, 180)]
    ])
    mask = np.zeros_like(r_image)
    # 绘制掩膜图像
    cv2.fillPoly(mask, poly, 255)
    # 获得ROI区域
    masked_image = cv2.bitwise_and(r_image, mask)
    return masked_image


if __name__ == '__main__':
    image = cv2.imread('test.jpg')
    lane_image = np.copy(image)
    canny = canny()
    cropped_image = region_of_interest(canny)
    cv2.imshow("result", cropped_image)
    cv2.waitKey(0)

霍夫变换加线性拟合改良：

效果图：

代码实现：

主要增加了根据斜率作线性拟合过滤无用点后连线的操作；

import cv2
import numpy as np


def canny():
    gray = cv2.cvtColor(lane_image, cv2.COLOR_RGB2GRAY)
    blur = cv2.GaussianBlur(gray, (5, 5), 0)

    canny_img = cv2.Canny(blur, 50, 150)
    return canny_img


def region_of_interest(r_image):
    h = r_image.shape[0]
    w = r_image.shape[1]

    poly = np.array([
        [(100, h), (500, h), (280, 180), (250, 180)]
    ])
    mask = np.zeros_like(r_image)
    cv2.fillPoly(mask, poly, 255)
    masked_image = cv2.bitwise_and(r_image, mask)
    return masked_image


def get_lines(img_lines):
    if img_lines is not None:
        for line in lines:
            for x1, y1, x2, y2 in line:
                # 分左右车道
                k = (y2 - y1) / (x2 - x1)
                if k < 0:
                    lefts.append(line)
                else:
                    rights.append(line)


def choose_lines(after_lines, slo_th):  # 过滤斜率差别较大的点
    slope = [(y2 - y1) / (x2 - x1) for line in after_lines for x1, x2, y1, y2 in line]  # 获得斜率数组
    while len(after_lines) > 0:
        mean = np.mean(slope)  # 计算平均斜率
        diff = [abs(s - mean) for s in slope]  # 每条线斜率与平均斜率的差距
        idx = np.argmax(diff)  # 找到最大斜率的索引
        if diff[idx] > slo_th:  # 大于预设的阈值选取
            slope.pop(idx)
            after_lines.pop(idx)
        else:
            break

    return after_lines


def clac_edgepoints(points, y_min, y_max):
    x = [p[0] for p in points]
    y = [p[1] for p in points]

    k = np.polyfit(y, x, 1)  # 曲线拟合的函数，找到xy的拟合关系斜率
    func = np.poly1d(k)  # 斜率代入可以得到一个y=kx的函数

    x_min = int(func(y_min))      # y_min = 325其实是近似找了一个
    x_max = int(func(y_max))

    return [(x_min, y_min), (x_max, y_max)]


if __name__ == '__main__':
    image = cv2.imread('F:\\A_javaPro\\test.jpg')
    lane_image = np.copy(image)
    canny_img = canny()
    cropped_image = region_of_interest(canny_img)
    lefts = []
    rights = []
    lines = cv2.HoughLinesP(cropped_image, 1, np.pi / 180, 15, np.array([]), minLineLength=40, maxLineGap=20)
    get_lines(lines)  # 分别得到左右车道线的图片

    good_leftlines = choose_lines(lefts, 0.1)  # 处理后的点
    good_rightlines = choose_lines(rights, 0.1)

    leftpoints = [(x1, y1) for left in good_leftlines for x1, y1, x2, y2 in left]
    leftpoints = leftpoints + [(x2, y2) for left in good_leftlines for x1, y1, x2, y2 in left]

    rightpoints = [(x1, y1) for right in good_rightlines for x1, y1, x2, y2 in right]
    rightpoints = rightpoints + [(x2, y2) for right in good_rightlines for x1, y1, x2, y2 in right]

    lefttop = clac_edgepoints(leftpoints, 180, image.shape[0])  # 要画左右车道线的端点
    righttop = clac_edgepoints(rightpoints, 180, image.shape[0])

    src = np.zeros_like(image)

    cv2.line(src, lefttop[0], lefttop[1], (255, 255, 0), 7)
    cv2.line(src, righttop[0], righttop[1], (255, 255, 0), 7)

    cv2.imshow('line Image', src)
    src_2 = cv2.addWeighted(image, 0.8, src, 1, 0)
    cv2.imshow('Finally Image', src_2)

    cv2.waitKey(0)

待改进：

代码实用性差，几乎不能用于实际，但是可以作为初学者的练手项目；
斑马线检测思路：获取车前感兴趣区域，判断白色像素点比例即可实现；
行人检测思路：opencv有内置行人检测函数，基于内置的训练好的数据集；