最近在unity仿真上做智能驾驶，车道线识别这块，网上找了很多教程，很多都行不通，要不就是不适用，总之各种问题，干脆自己写一个（参考了很多前人的经验），代码粘在最下面，备忘。不bb，先上图

首先是原图

然后是效果图

unity截图

unity效果截图

一、整体思路：

opencv：

看起来复杂，其实套路都一样，就是调参的时候费劲。

 我是基于图片写的，视频的话需要拆成一帧一帧的，原理同下。

1、首先高斯降噪。GaussianBlur（）

2、然后，灰度，二值，边缘检测。  cvtColor->threshold->filter2D

    这里推荐边缘检测用卷积，其他的方式也行，但是感觉效果不太好。

3、扣下敏感部分，画框填充，&抠图。fillConvexPoly->bitwise_and

4、重点，霍夫曼直线检测。HoughLinesP，

   概率Hough变换，HoughLinesP只通过分析点的子集并估计这些点都属于一条直线，说白了是因为有的车道线是虚线，所以不推荐用HoughLines

5、将所有线分成左线右线。首先根据直线斜率找出直线，然后根据x轴分成左线右线

6、对根据得到的信息进行直线拟合，fitLine，得到四个顶点

7、根据直线斜率判断行驶方向。

8、将线画在图上。

unity：

unity仿真端就是将相机得到的图片保存下来，然后调用c++打包的dll。

首先，放个摄像机，给摄像机一个texture

然后，将这个texture生成到本地。

代码见下

还有一点很重要的是，我的方法是将unity的图片生成到本地，也就是外存，然后opencv读图图片，再处理，如果想将图片显示在unity界面上，opencv还要将图片生成到本地，unity再读取加载，光是内外存速度问题，这样一大圈下来，也会让程序非常非常非常非常非常卡，目前有两种解决办法，

该问题已解决：详情见另一篇博文，实现了unity3d和opencv的实时高效图像传递，跳转

再贴一遍地址：https://blog.csdn.net/qq_25490573/article/details/99671398

第一种是，unity将相机的图片生成在内存里，然后让opencv的dll直接在内存直接进行操作。

        这样一来就有很多问题了，考虑到两种语言对照问题，用了byte和无符号char，首先，需要两边生成图片的格式一样，比如opencv解图片是三通道，unity需要对应rgb24，但实际解的时候问题频出，unity说崩就崩。unity图片0点是左下角，opencv却是右上角。这意味着现有的Encode...方法的不适用。

       解决这个问题的方法是。。。项目被领导砍了，，，有生之年再补上

另一种方法是，借助opencvforunity插件实现，实际也是上面的方法，只是人家给你封装好了不需要你再进行操作。

二、代码实现

这个头文件是必须的

#include <string>
#include <vector>
#include <opencv2/opencv.hpp>

class LaneDetector
{
private:
	double img_size;
	double img_center;
	bool left_flag = false;  
	bool right_flag = false;  
	cv::Point right_b;  
	double right_m;  // y = m*x + b
	cv::Point left_b; 
	double left_m; 

public:
	cv::Mat deNoise(cv::Mat inputImage);  // Apply Gaussian blurring to the input Image
	cv::Mat edgeDetector(cv::Mat img_noise);  // Filter the image to obtain only edges
	cv::Mat mask(cv::Mat img_edges);  // Mask the edges image to only care about ROI
	std::vector<cv::Vec4i> houghLines(cv::Mat img_mask);  // Detect Hough lines in masked edges image
	std::vector<std::vector<cv::Vec4i> > lineSeparation(std::vector<cv::Vec4i> lines, cv::Mat img_edges);  // Sprt detected lines by their slope into right and left lines
	std::vector<cv::Point> regression(std::vector<std::vector<cv::Vec4i> > left_right_lines, cv::Mat inputImage);  // Get only one line for each side of the lane
	std::string predictTurn();  // Determine if the lane is turning or not by calculating the position of the vanishing point
	int plotLane(cv::Mat inputImage, std::vector<cv::Point> lane, std::string turn);  // Plot the resultant lane and turn prediction in the frame.
};
//extern "C"  __declspec(dllexport) int lanedet();

这个cpp也是必须要有的

#include "stdafx.h"
#include <string>
#include <vector>
#include <opencv2/opencv.hpp>
#include "LaneDetector.h"

// IMAGE BLURRING
/**
*@brief Apply gaussian filter to the input image to denoise it
*@param inputImage is the frame of a video in which the
*@param lane is going to be detected
*@return Blurred and denoised image
*/
cv::Mat LaneDetector::deNoise(cv::Mat inputImage) {
	cv::Mat output;

	cv::GaussianBlur(inputImage, output, cv::Size(3, 3), 0, 0);

	return output;
}

// EDGE DETECTION
/**
*@brief Detect all the edges in the blurred frame by filtering the image
*@param img_noise is the previously blurred frame
*@return Binary image with only the edges represented in white
*/
cv::Mat LaneDetector::edgeDetector(cv::Mat img_noise) {
	cv::Mat output;
	cv::Mat kernel;
	cv::Point anchor;

	// Convert image from RGB to gray
	cv::cvtColor(img_noise, output, cv::COLOR_RGB2GRAY);
	// Binarize gray image
	cv::threshold(output, output, 140, 255, cv::THRESH_BINARY);

	// Create the kernel [-1 0 1]
	// This kernel is based on the one found in the
	// Lane Departure Warning System by Mathworks
	anchor = cv::Point(-1, -1);
	kernel = cv::Mat(1, 3, CV_32F);
	kernel.at<float>(0, 0) = -1;
	kernel.at<float>(0, 1) = 0;
	kernel.at<float>(0, 2) = 1;

	// Filter the binary image to obtain the edges
	cv::filter2D(output, output, -1, kernel, anchor, 0, cv::BORDER_DEFAULT);
	cv::imshow("output", output);
	return output;
}

// MASK THE EDGE IMAGE
/**
*@brief Mask the image so that only the edges that form part of the lane are detected
*@param img_edges is the edges image from the previous function
*@return Binary image with only the desired edges being represented
*/
cv::Mat LaneDetector::mask(cv::Mat img_edges) {
	cv::Mat output;
	cv::Mat mask = cv::Mat::zeros(img_edges.size(), img_edges.type());
	cv::Point pts[4] = {
		cv::Point(210, 720),
		cv::Point(550, 450),
		cv::Point(717, 450),
		cv::Point(1280, 720)
	};

	// Create a binary polygon mask
	cv::fillConvexPoly(mask, pts, 4, cv::Scalar(255, 0, 0));
	// Multiply the edges image and the mask to get the output
	cv::bitwise_and(img_edges, mask, output);

	return output;
}

// HOUGH LINES
/**
*@brief Obtain all the line segments in the masked images which are going to be part of the lane boundaries
*@param img_mask is the masked binary image from the previous function
*@return Vector that contains all the detected lines in the image
*/
std::vector<cv::Vec4i> LaneDetector::houghLines(cv::Mat img_mask) {
	std::vector<cv::Vec4i> line;

	// rho and theta are selected by trial and error
	HoughLinesP(img_mask, line, 1, CV_PI / 180, 20, 20, 30);

	return line;
}

// SORT RIGHT AND LEFT LINES
/**
*@brief Sort all the detected Hough lines by slope.
*@brief The lines are classified into right or left depending
*@brief on the sign of their slope and their approximate location
*@param lines is the vector that contains all the detected lines
*@param img_edges is used for determining the image center
*@return The output is a vector(2) that contains all the classified lines
*/
std::vector<std::vector<cv::Vec4i> > LaneDetector::lineSeparation(std::vector<cv::Vec4i> lines, cv::Mat img_edges) {
	std::vector<std::vector<cv::Vec4i> > output(2);
	size_t j = 0;
	cv::Point ini;
	cv::Point fini;
	double slope_thresh = 0.3;
	std::vector<double> slopes;
	std::vector<cv::Vec4i> selected_lines;
	std::vector<cv::Vec4i> right_lines, left_lines;

	// Calculate the slope of all the detected lines
	for (auto i : lines) {
		ini = cv::Point(i[0], i[1]);
		fini = cv::Point(i[2], i[3]);

		// Basic algebra: slope = (y1 - y0)/(x1 - x0)
		double slope = (static_cast<double>(fini.y) - static_cast<double>(ini.y)) / (static_cast<double>(fini.x) - static_cast<double>(ini.x) + 0.00001);

		// If the slope is too horizontal, discard the line
		// If not, save them  and their respective slope
		if (std::abs(slope) > slope_thresh) {
			slopes.push_back(slope);
			selected_lines.push_back(i);
		}
	}

	// Split the lines into right and left lines
	img_center = static_cast<double>((img_edges.cols / 2));
	while (j < selected_lines.size()) {
		ini = cv::Point(selected_lines[j][0], selected_lines[j][1]);
		fini = cv::Point(selected_lines[j][2], selected_lines[j][3]);

		// Condition to classify line as left side or right side
		if (slopes[j] > 0 && fini.x > img_center && ini.x > img_center) {
			right_lines.push_back(selected_lines[j]);
			right_flag = true;
		}
		else if (slopes[j] < 0 && fini.x < img_center && ini.x < img_center) {
			left_lines.push_back(selected_lines[j]);
			left_flag = true;
		}
		j++;
	}

	output[0] = right_lines;
	output[1] = left_lines;

	return output;
}

// REGRESSION FOR LEFT AND RIGHT LINES
/**
*@brief Regression takes all the classified line segments initial and final points and fits a new lines out of them using the method of least squares.
*@brief This is done for both sides, left and right.
*@param left_right_lines is the output of the lineSeparation function
*@param inputImage is used to select where do the lines will end
*@return output contains the initial and final points of both lane boundary lines
*/
std::vector<cv::Point> LaneDetector::regression(std::vector<std::vector<cv::Vec4i> > left_right_lines, cv::Mat inputImage) {
	std::vector<cv::Point> output(4);
	cv::Point ini;
	cv::Point fini;
	cv::Point ini2;
	cv::Point fini2;
	cv::Vec4d right_line;
	cv::Vec4d left_line;
	std::vector<cv::Point> right_pts;
	std::vector<cv::Point> left_pts;

	// If right lines are being detected, fit a line using all the init and final points of the lines
	if (right_flag == true) {
		for (auto i : left_right_lines[0]) {
			ini = cv::Point(i[0], i[1]);
			fini = cv::Point(i[2], i[3]);

			right_pts.push_back(ini);
			right_pts.push_back(fini);
		}

		if (right_pts.size() > 0) {
			// The right line is formed here
			cv::fitLine(right_pts, right_line, CV_DIST_L2, 0, 0.01, 0.01);
			right_m = right_line[1] / right_line[0];
			right_b = cv::Point(right_line[2], right_line[3]);
		}
	}

	// If left lines are being detected, fit a line using all the init and final points of the lines
	if (left_flag == true) {
		for (auto j : left_right_lines[1]) {
			ini2 = cv::Point(j[0], j[1]);
			fini2 = cv::Point(j[2], j[3]);

			left_pts.push_back(ini2);
			left_pts.push_back(fini2);
		}

		if (left_pts.size() > 0) {
			// The left line is formed here
			cv::fitLine(left_pts, left_line, CV_DIST_L2, 0, 0.01, 0.01);
			left_m = left_line[1] / left_line[0];
			left_b = cv::Point(left_line[2], left_line[3]);
		}
	}

	// One the slope and offset points have been obtained, apply the line equation to obtain the line points
	int ini_y = inputImage.rows;
	int fin_y = 470;

	double right_ini_x = ((ini_y - right_b.y) / right_m) + right_b.x;
	double right_fin_x = ((fin_y - right_b.y) / right_m) + right_b.x;

	double left_ini_x = ((ini_y - left_b.y) / left_m) + left_b.x;
	double left_fin_x = ((fin_y - left_b.y) / left_m) + left_b.x;

	output[0] = cv::Point(right_ini_x, ini_y);
	output[1] = cv::Point(right_fin_x, fin_y);
	output[2] = cv::Point(left_ini_x, ini_y);
	output[3] = cv::Point(left_fin_x, fin_y);

	return output;
}

// TURN PREDICTION
/**
*@brief Predict if the lane is turning left, right or if it is going straight
*@brief It is done by seeing where the vanishing point is with respect to the center of the image
*@return String that says if there is left or right turn or if the road is straight
*/
std::string LaneDetector::predictTurn() {
	std::string output;
	double vanish_x;
	double thr_vp = 10;

	// The vanishing point is the point where both lane boundary lines intersect
	vanish_x = static_cast<double>(((right_m*right_b.x) - (left_m*left_b.x) - right_b.y + left_b.y) / (right_m - left_m));

	// The vanishing points location determines where is the road turning
	if (vanish_x < (img_center - thr_vp))
		output = "Left Turn";
	else if (vanish_x >(img_center + thr_vp))
		output = "Right Turn";
	else if (vanish_x >= (img_center - thr_vp) && vanish_x <= (img_center + thr_vp))
		output = "Straight";

	return output;
}

// PLOT RESULTS
/**
*@brief This function plots both sides of the lane, the turn prediction message and a transparent polygon that covers the area inside the lane boundaries
*@param inputImage is the original captured frame
*@param lane is the vector containing the information of both lines
*@param turn is the output string containing the turn information
*@return The function returns a 0
*/
int LaneDetector::plotLane(cv::Mat inputImage, std::vector<cv::Point> lane, std::string turn) {
	std::vector<cv::Point> poly_points;
	cv::Mat output;

	// Create the transparent polygon for a better visualization of the lane
	inputImage.copyTo(output);
	poly_points.push_back(lane[2]);
	poly_points.push_back(lane[0]);
	poly_points.push_back(lane[1]);
	poly_points.push_back(lane[3]);
	cv::fillConvexPoly(output, poly_points, cv::Scalar(0, 0, 255), CV_AA, 0);
	cv::addWeighted(output, 0.3, inputImage, 1.0 - 0.3, 0, inputImage);

	// Plot both lines of the lane boundary
	cv::line(inputImage, lane[0], lane[1], cv::Scalar(0, 255, 255), 5, CV_AA);
	cv::line(inputImage, lane[2], lane[3], cv::Scalar(0, 255, 255), 5, CV_AA);

	// Plot the turn message
	cv::putText(inputImage, turn, cv::Point(50, 90), cv::FONT_HERSHEY_COMPLEX_SMALL, 3, cvScalar(0, 255, 0), 1, CV_AA);

	// Show the final output image
	cv::namedWindow("Lane", CV_WINDOW_AUTOSIZE);
	cv::imshow("Lane", inputImage);
	return 0;
}

如果播放的视频，用这个cpp

#include "stdafx.h"
#include <iostream>
#include <string>
#include <vector>
#include <opencv2/opencv.hpp>
#include "LaneDetector.h"
//#include "LaneDetector.cpp"

/**
*@brief Function main that runs the main algorithm of the lane detection.
*@brief It will read a video of a car in the highway and it will output the
*@brief same video but with the plotted detected lane
*@param argv[] is a string to the full path of the demo video
*@return flag_plot tells if the demo has sucessfully finished
*/
int main() {

	// The input argument is the location of the video
	cv::VideoCapture cap("project_video.mp4");
	if (!cap.isOpened())
		return -1;

	LaneDetector lanedetector;  // Create the class object
	cv::Mat frame;
	cv::Mat img_denoise;
	cv::Mat img_edges;
	cv::Mat img_mask;
	cv::Mat img_lines;
	std::vector<cv::Vec4i> lines;
	std::vector<std::vector<cv::Vec4i> > left_right_lines;
	std::vector<cv::Point> lane;
	std::string turn;
	int flag_plot = -1;
	int i = 0;
	// Main algorithm starts. Iterate through every frame of the video
	while (i < 540) {
		// Capture frame
		if (!cap.read(frame))
			break;

		// Denoise the image using a Gaussian filter
		img_denoise = lanedetector.deNoise(frame);

		// Detect edges in the image
		img_edges = lanedetector.edgeDetector(img_denoise);

		// Mask the image so that we only get the ROI
		img_mask = lanedetector.mask(img_edges);

		// Obtain Hough lines in the cropped image
		lines = lanedetector.houghLines(img_mask);

		if (!lines.empty())
		{
			// Separate lines into left and right lines
			left_right_lines = lanedetector.lineSeparation(lines, img_edges);

			// Apply regression to obtain only one line for each side of the lane
			lane = lanedetector.regression(left_right_lines, frame);

			// Predict the turn by determining the vanishing point of the the lines
			turn = lanedetector.predictTurn();

			// Plot lane detection
			flag_plot = lanedetector.plotLane(frame, lane, turn);

			i += 1;
			cv::waitKey(25);
		}
		else {
			flag_plot = -1;
		}
	}
	
	return flag_plot;
}

unity的dll打包用这个，如果需要的是控制台程序，只需要将lanedet函数换成main函数

打包dll需要在第一个脚本的最后一行添加这句话才能用

extern "C"  __declspec(dllexport) int lanedet();

#include "stdafx.h"
#include <iostream>
#include <string>
#include <vector>
#include <opencv2/opencv.hpp>
#include "LaneDetector.h"
//using namespace cv;
//using namespace std;

LaneDetector lanedetector;  // Create the class object
cv::Mat frame;
cv::Mat img_denoise;
cv::Mat img_edges;
cv::Mat img_mask;
cv::Mat img_lines;
std::vector<cv::Vec4i> lines;
std::vector<std::vector<cv::Vec4i> > left_right_lines;
std::vector<cv::Point> lane;
std::string turn;
int flag_plot = -1;
int i = 0;

int lanedet(){

	//int lanedet(BYTE * src)
	frame = cv::imread("D:\\hehe.png"); //cv::imread("C:\\Users\\Administrator\\Desktop\\image\\haha.png");
	//imshow("asdf",frame);
	//cv::waitKey();
	
	img_denoise = lanedetector.deNoise(frame);

	// Detect edges in the image
	img_edges = lanedetector.edgeDetector(img_denoise);

	// Mask the image so that we only get the ROI
	img_mask = lanedetector.mask(img_edges);

	// Obtain Hough lines in the cropped image
	lines = lanedetector.houghLines(img_mask);

	if (!lines.empty())
	{
		// Separate lines into left and right lines
		left_right_lines = lanedetector.lineSeparation(lines, img_edges);

		// Apply regression to obtain only one line for each side of the lane
		lane = lanedetector.regression(left_right_lines, frame);

		// Predict the turn by determining the vanishing point of the the lines
		turn = lanedetector.predictTurn();

		// Plot lane detection
		flag_plot = lanedetector.plotLane(frame, lane, turn);

		//i += 1;
		cv::waitKey(20);
	}
	else {
		//flag_plot = -1;
		return 0;
	}

	getchar();
	return 1;
}

unity生成图片方法和调用dll的方法

using System.Collections;
using System.Collections.Generic;
using UnityEngine;
using System.IO;
using System.Runtime.InteropServices;
using System.Threading;



public class LaneDetector : MonoBehaviour {

    [DllImport("LaneIdentificationImg")]
    extern static int lanedet();

    [SerializeField]
    Camera sensorcamera;
    RenderTexture rt;

    

    void Start () {
        InvokeRepeating("imginvoke",5,1f);

    }
	
	void Update () {
        rt = sensorcamera.targetTexture;
    }
    void imginvoke()
    {
        OutputRt(sensorcamera.targetTexture, 0);
        lanedet();
    }

    public void OutputRt(RenderTexture rt, int idx = 0)
    {
        rt = sensorcamera.targetTexture;
        RenderTexture.active = rt;

        Texture2D png = new Texture2D(rt.width, rt.height, TextureFormat.RGB24, false);
        png.ReadPixels(new Rect(0, 0, rt.width, rt.height), 0, 0);
      
        byte[] dataBytes = png.EncodeToPNG();
        //int[] data = new int[dataBytes.Length];
        //for (int i = 0; i < dataBytes.Length; i++)
        //{
        //    data[i] = (int)dataBytes[i];
        //}
        //if (lanedet(dataBytes) == 0)
        //{
        //    return;
        //}
        string strSaveFile = "D:\\hehe" + ".png";
        FileStream fs = File.Open(strSaveFile, FileMode.OpenOrCreate);
        fs.Write(dataBytes, 0, dataBytes.Length);
        fs.Flush();
        fs.Close();
        png = null;
        RenderTexture.active = null;
    }
}