用opencvSharp实现最小二乘法和Ransac随机抽样一致法来拟合直线

            var surf = OpenCvSharp.XFeatures2D.SURF.Create(2000);
            var keyPoints = surf.Detect(src);
            Cv2.DrawKeypoints(src, keyPoints, src);
            surf.Dispose();

            Vector2[] vector2s = new Vector2[keyPoints.Length];
            Point[] points = new Point[keyPoints.Length];
            for (int i = 0; i < keyPoints.Length; i++)
            {
                var point = (Point)keyPoints[i].Pt;
                points[i] = point;
                vector2s[i].X = point.X;
                vector2s[i].Y = point.Y;
            }

最小二乘法在opencvsharp里有现成的api调用，这里就不多说了，很简单：

            var line = Cv2.FitLine(points, DistanceTypes.L1, 0, 0, 0);
            line.FitSize(src.Width, src.Height, out var p1, out var p2);
            Cv2.Line(src, p1, p2, Scalar.Blue, 5);

然后是Ransac随机抽样一致法，和网络上的办法一样，也是用距离来判断点是否在给定阈值范围内。
结果图，红色的线是Ransac，蓝色的线是最小二乘法：

完整代码：

        public Form1()
        {
            InitializeComponent();
            var src = Cv2.ImRead("C:\\Users\\Administrator\\Desktop\\test2.png", ImreadModes.Grayscale);
            TestFitLine(src);
            Cv2.ImShow("dst", src);
        }

        public static void TestFitLine(Mat src)
        {
            var surf = OpenCvSharp.XFeatures2D.SURF.Create(2000);
            var keyPoints = surf.Detect(src);
            Cv2.DrawKeypoints(src, keyPoints, src);
            surf.Dispose();

            Vector2[] vector2s = new Vector2[keyPoints.Length];
            Point[] points = new Point[keyPoints.Length];
            for (int i = 0; i < keyPoints.Length; i++)
            {
                var point = (Point)keyPoints[i].Pt;
                points[i] = point;
                vector2s[i].X = point.X;
                vector2s[i].Y = point.Y;
            }

            //最小二乘法
            var line = Cv2.FitLine(points, DistanceTypes.L1, 0, 0, 0);
            line.FitSize(src.Width, src.Height, out var p1, out var p2);
            Cv2.Line(src, p1, p2, Scalar.Blue, 5);

            //Ransac随机抽样一致法
            var r = FitLineRansac(vector2s, 1000, 10);
            r.FitSize(src.Width, src.Height, out var p3, out var p4);
            Cv2.Line(src, p3, p4, Scalar.Red, 5);
        }

        public static Line2D FitLineRansac(Vector2[] points, int iterations, double sigma)
        {
            int bestScore = -1;
            Line2D result = null;
            Random random = new Random();
            for (int i = 0; i < iterations; i++)
            {
                if (bestScore > points.Length * 0.5)
                    break;

                var indexes = GetRandomIndexes(random, 0, points.Length);
                var p1 = points[indexes.Item0];
                var p2 = points[indexes.Item1];

                var dir = Vector2.Normalize(p2 - p1);
                var line = new Line2D(dir.X, dir.Y, p2.X, p2.Y);
                int score = 0;
                for (int j = 0; j < points.Length; j++)
                {
                    if (line.Distance(points[j].X, points[j].Y) < sigma)
                        score += 1;
                }

                if (score > bestScore)
                {
                    bestScore = score;
                    result = line;
                }
            }
            return result;
        }

        private static Vec2i GetRandomIndexes(Random random, int min, int max)
        {
            var index1 = random.Next(min, max);
            var index2 = random.Next(min, max);
            if (index1 == index2)
                GetRandomIndexes(random, min, max);
            return new Vec2i(index1, index2);
        }