nirg回答的Swift版本:

extension CGPoint {
    func isInsidePolygon(vertices: [CGPoint]) -> Bool {
        guard !vertices.isEmpty else { return false }
        var j = vertices.last!, c = false
        for i in vertices {
            let a = (i.y > y) != (j.y > y)
            let b = (x < (j.x - i.x) * (y - i.y) / (j.y - i.y) + i.x)
            if a && b { c = !c }
            j = i
        }
        return c
    }
}

For graphics, I'd rather not prefer integers. Many systems use integers for UI painting (pixels are ints after all), but macOS, for example, uses float for everything. macOS only knows points and a point can translate to one pixel, but depending on monitor resolution, it might translate to something else. On retina screens half a point (0.5/0.5) is pixel. Still, I never noticed that macOS UIs are significantly slower than other UIs. After all, 3D APIs (OpenGL or Direct3D) also work with floats and modern graphics libraries very often take advantage of GPU acceleration.

现在你说速度是你最关心的，好吧，让我们追求速度。在运行任何复杂的算法之前，先做一个简单的测试。在多边形周围创建一个轴对齐的包围框。这是非常简单，快速的，已经可以节省你很多计算。这是怎么做到的呢?遍历多边形的所有点，找到X和Y的最小/最大值。

如你有点(9/1),(4/3),(2/7),(8/2),(3/6)。这意味着Xmin是2,Xmax是9,Ymin是1,Ymax是7。矩形外有两条边(2/1)和(9/7)的点不可能在多边形内。

// p is your point, p.x is the x coord, p.y is the y coord
if (p.x < Xmin || p.x > Xmax || p.y < Ymin || p.y > Ymax) {
    // Definitely not within the polygon!
}

这是对任意点运行的第一个测试。正如你所看到的，这个测试非常快，但也非常粗糙。要处理边界矩形内的点，我们需要更复杂的算法。有几种计算方法。哪种方法有效还取决于多边形是否有孔或始终是固体。以下是实体的例子(一个凸面，一个凹面):

这里有一个洞:

绿色的中间有个洞!

最简单的算法，可以处理上述三种情况，并且仍然非常快，叫做射线投射。该算法的思想非常简单:从多边形外的任何地方绘制一条虚拟光线到你的点，并计算它击中多边形一侧的频率。如果命中次数是偶数，则在多边形外，如果是奇数，则在多边形内。

圈数算法是另一种选择，它对非常接近多边形线的点更准确，但也慢得多。由于有限的浮点精度和舍入问题，光线投射可能会因为太靠近多边形一侧的点而失败，但在现实中这几乎不是问题，因为如果一个点靠近一侧，在视觉上甚至不可能让观看者识别它是否已经在内部或仍然在外部。

还记得上面的边界框吗?只需在边界框外选择一个点，并将其用作射线的起点。例如，点(Xmin - e/p.y)肯定在多边形外。

But what is e? Well, e (actually epsilon) gives the bounding box some padding. As I said, ray tracing fails if we start too close to a polygon line. Since the bounding box might equal the polygon (if the polygon is an axis aligned rectangle, the bounding box is equal to the polygon itself!), we need some padding to make this safe, that's all. How big should you choose e? Not too big. It depends on the coordinate system scale you use for drawing. If your pixel step width is 1.0, then just choose 1.0 (yet 0.1 would have worked as well)

现在我们有了光线的起始坐标和结束坐标，问题从“是多边形内的点”转移到“光线与多边形边相交的频率”。因此，我们不能像以前那样只处理多边形点，现在我们需要实际的边。一条边总是由两点来定义。

side 1: (X1/Y1)-(X2/Y2)
side 2: (X2/Y2)-(X3/Y3)
side 3: (X3/Y3)-(X4/Y4)
:

你需要从各个方面测试光线。假设射线是一个矢量，每条边都是一个矢量。光线必须恰好击中每边一次，否则就永远不会。它不可能击中同一侧两次。二维空间中的两条直线总是只相交一次，除非它们是平行的，在这种情况下，它们永远不会相交。然而，由于向量的长度是有限的，两个向量可能不平行，也永远不会相交，因为它们太短而无法相遇。

// Test the ray against all sides
int intersections = 0;
for (side = 0; side < numberOfSides; side++) {
    // Test if current side intersects with ray.
    // If yes, intersections++;
}
if ((intersections & 1) == 1) {
    // Inside of polygon
} else {
    // Outside of polygon
}

到目前为止一切顺利，但是如何检验两个向量是否相交呢?下面是一些C代码(未测试)，应该可以做到:

#define NO 0
#define YES 1
#define COLLINEAR 2

int areIntersecting(
    float v1x1, float v1y1, float v1x2, float v1y2,
    float v2x1, float v2y1, float v2x2, float v2y2
) {
    float d1, d2;
    float a1, a2, b1, b2, c1, c2;

    // Convert vector 1 to a line (line 1) of infinite length.
    // We want the line in linear equation standard form: A*x + B*y + C = 0
    // See: http://en.wikipedia.org/wiki/Linear_equation
    a1 = v1y2 - v1y1;
    b1 = v1x1 - v1x2;
    c1 = (v1x2 * v1y1) - (v1x1 * v1y2);

    // Every point (x,y), that solves the equation above, is on the line,
    // every point that does not solve it, is not. The equation will have a
    // positive result if it is on one side of the line and a negative one 
    // if is on the other side of it. We insert (x1,y1) and (x2,y2) of vector
    // 2 into the equation above.
    d1 = (a1 * v2x1) + (b1 * v2y1) + c1;
    d2 = (a1 * v2x2) + (b1 * v2y2) + c1;

    // If d1 and d2 both have the same sign, they are both on the same side
    // of our line 1 and in that case no intersection is possible. Careful, 
    // 0 is a special case, that's why we don't test ">=" and "<=", 
    // but "<" and ">".
    if (d1 > 0 && d2 > 0) return NO;
    if (d1 < 0 && d2 < 0) return NO;

    // The fact that vector 2 intersected the infinite line 1 above doesn't 
    // mean it also intersects the vector 1. Vector 1 is only a subset of that
    // infinite line 1, so it may have intersected that line before the vector
    // started or after it ended. To know for sure, we have to repeat the
    // the same test the other way round. We start by calculating the 
    // infinite line 2 in linear equation standard form.
    a2 = v2y2 - v2y1;
    b2 = v2x1 - v2x2;
    c2 = (v2x2 * v2y1) - (v2x1 * v2y2);

    // Calculate d1 and d2 again, this time using points of vector 1.
    d1 = (a2 * v1x1) + (b2 * v1y1) + c2;
    d2 = (a2 * v1x2) + (b2 * v1y2) + c2;

    // Again, if both have the same sign (and neither one is 0),
    // no intersection is possible.
    if (d1 > 0 && d2 > 0) return NO;
    if (d1 < 0 && d2 < 0) return NO;

    // If we get here, only two possibilities are left. Either the two
    // vectors intersect in exactly one point or they are collinear, which
    // means they intersect in any number of points from zero to infinite.
    if ((a1 * b2) - (a2 * b1) == 0.0f) return COLLINEAR;

    // If they are not collinear, they must intersect in exactly one point.
    return YES;
}

输入值是向量1 (v1x1/v1y1和v1x2/v1y2)和向量2 (v2x1/v2y1和v2x2/v2y2)的两个端点。有2个向量，4个点，8个坐标。YES和NO很清楚。“是”增加了交叉路口，“否”什么都不做。

What about COLLINEAR? It means both vectors lie on the same infinite line, depending on position and length, they don't intersect at all or they intersect in an endless number of points. I'm not absolutely sure how to handle this case, I would not count it as intersection either way. Well, this case is rather rare in practice anyway because of floating point rounding errors; better code would probably not test for == 0.0f but instead for something like < epsilon, where epsilon is a rather small number.

如果你需要测试更多的点，你当然可以通过在内存中保留多边形边的线性方程标准形式来加快整个过程，这样你就不必每次都重新计算这些点。这将在每次测试中为您节省两次浮点乘法和三次浮点减法，以换取在内存中为每个多边形边存储三个浮点值。这是一个典型的内存与计算时间的权衡。

Last but not least: If you may use 3D hardware to solve the problem, there is an interesting alternative. Just let the GPU do all the work for you. Create a painting surface that is off screen. Fill it completely with the color black. Now let OpenGL or Direct3D paint your polygon (or even all of your polygons if you just want to test if the point is within any of them, but you don't care for which one) and fill the polygon(s) with a different color, e.g. white. To check if a point is within the polygon, get the color of this point from the drawing surface. This is just a O(1) memory fetch.

Of course this method is only usable if your drawing surface doesn't have to be huge. If it cannot fit into the GPU memory, this method is slower than doing it on the CPU. If it would have to be huge and your GPU supports modern shaders, you can still use the GPU by implementing the ray casting shown above as a GPU shader, which absolutely is possible. For a larger number of polygons or a large number of points to test, this will pay off, consider some GPUs will be able to test 64 to 256 points in parallel. Note however that transferring data from CPU to GPU and back is always expensive, so for just testing a couple of points against a couple of simple polygons, where either the points or the polygons are dynamic and will change frequently, a GPU approach will rarely pay off.

没有什么比归纳定义问题更美好的了。为了完整起见，你在序言中有一个版本，它可能也澄清了光线投射背后的思想:

基于仿真的简化算法在http://www.ecse.rpi.edu/Homepages/wrf/Research/Short_Notes/pnpoly.html

一些helper谓词:

exor(A,B):- \+A,B;A,\+B.
in_range(Coordinate,CA,CB) :- exor((CA>Coordinate),(CB>Coordinate)).

inside(false).
inside(_,[_|[]]).
inside(X:Y, [X1:Y1,X2:Y2|R]) :- in_range(Y,Y1,Y2), X > ( ((X2-X1)*(Y-Y1))/(Y2-Y1) +      X1),toggle_ray, inside(X:Y, [X2:Y2|R]); inside(X:Y, [X2:Y2|R]).

get_line(_,_,[]).
get_line([XA:YA,XB:YB],[X1:Y1,X2:Y2|R]):- [XA:YA,XB:YB]=[X1:Y1,X2:Y2]; get_line([XA:YA,XB:YB],[X2:Y2|R]).

给定两点a和B的直线(直线(a,B))方程为:

                    (YB-YA)
           Y - YA = ------- * (X - XA) 
                    (XB-YB) 

It is important that the direction of rotation for the line is setted to clock-wise for boundaries and anti-clock-wise for holes. We are going to check whether the point (X,Y), i.e the tested point is at the left half-plane of our line (it is a matter of taste, it could also be the right side, but also the direction of boundaries lines has to be changed in that case), this is to project the ray from the point to the right (or left) and acknowledge the intersection with the line. We have chosen to project the ray in the horizontal direction (again it is a matter of taste, it could also be done in vertical with similar restrictions), so we have:

               (XB-XA)
           X < ------- * (Y - YA) + XA
               (YB-YA) 

Now we need to know if the point is at the left (or right) side of the line segment only, not the entire plane, so we need to restrict the search only to this segment, but this is easy since to be inside the segment only one point in the line can be higher than Y in the vertical axis. As this is a stronger restriction it needs to be the first to check, so we take first only those lines meeting this requirement and then check its possition. By the Jordan Curve theorem any ray projected to a polygon must intersect at an even number of lines. So we are done, we will throw the ray to the right and then everytime it intersects a line, toggle its state. However in our implementation we are goint to check the lenght of the bag of solutions meeting the given restrictions and decide the innership upon it. for each line in the polygon this have to be done.

is_left_half_plane(_,[],[],_).
is_left_half_plane(X:Y,[XA:YA,XB:YB], [[X1:Y1,X2:Y2]|R], Test) :- [XA:YA, XB:YB] =  [X1:Y1, X2:Y2], call(Test, X , (((XB - XA) * (Y - YA)) / (YB - YA) + XA)); 
                                                        is_left_half_plane(X:Y, [XA:YA, XB:YB], R, Test).

in_y_range_at_poly(Y,[XA:YA,XB:YB],Polygon) :- get_line([XA:YA,XB:YB],Polygon),  in_range(Y,YA,YB).
all_in_range(Coordinate,Polygon,Lines) :- aggregate(bag(Line),    in_y_range_at_poly(Coordinate,Line,Polygon), Lines).

traverses_ray(X:Y, Lines, Count) :- aggregate(bag(Line), is_left_half_plane(X:Y, Line, Lines, <), IntersectingLines), length(IntersectingLines, Count).

% This is the entry point predicate
inside_poly(X:Y,Polygon,Answer) :- all_in_range(Y,Polygon,Lines), traverses_ray(X:Y, Lines, Count), (1 is mod(Count,2)->Answer=inside;Answer=outside).

在大多数情况下，这是一个比其他算法都快的算法。

它又新又雅致。我们花费O(n * log(n))时间构建一个表，允许我们在O(log(n) + k)时间内测试多边形中的点。

与光线跟踪或角度不同，使用扫描光束表可以更快地对同一多边形进行多次检查。我们必须预先构建一个扫描束活动边表，这是大多数代码正在做的事情。

We calculate the scanbeam and the active edges for that position in the y-direction. We make a list of points sorted by their y-component and we iterate through this list, for two events. Start-Y and End-Y, we track the active edges as we process the list. We process the events in order and for each scanbeam we record the y-value of the event and the active edges at each event (events being start-y and end-y) but we only record these when our event-y is different than last time (so everything at the event point is processed before we mark it in our table).

我们得到我们的表格:

[] p6p5、p6p7 p6p5, p6p7, p2p3, p2p1 p6p7, p5p4, p2p3, p3p1 p7p8, p5p4, p2p3, p2p1 p7p8, p5p4, p3p4, p2p1 p7p8 p2p1、 p7p8、p1p0 p8p0、p1p0 []

在构建该表之后，实际执行工作的代码只有几行。

注意:这里的代码使用复数值作为点。所以。real是。x。imag是。y。

def point_in_scantable(actives_table, events, xi, point):
    beam = bisect(events, point.imag) - 1  # Binary search in sorted array.
    actives_at_y = actives_table[beam]
    total = sum([point.real > xi(e, point.imag) for e in actives_at_y])
    return bool(total % 2)

我们对事件进行二进制搜索，以找到特定值的actives_at_y。对于在y点的所有活动，我们计算在我们点的特定y点的x段值。每次x截距大于点的x分量时加1。然后对总数乘以2。(这是偶数-奇数填充规则，你可以很容易地适应任何其他填充规则)。

完整的代码:

from bisect import bisect

def build_edge_list(polygon):
    edge_list = []
    for i in range(1, len(polygon)):
        if (polygon[i].imag, polygon[i].real) < (polygon[i - 1].imag, polygon[i - 1].real):
            edge_list.append((polygon[i], i))
            edge_list.append((polygon[i - 1], ~i))
        else:
            edge_list.append((polygon[i], ~i))
            edge_list.append((polygon[i - 1], i))

    def sort_key(e):
        return e[0].imag, e[0].real, ~e[1]

    edge_list.sort(key=sort_key)
    return edge_list


def build_scanbeam(edge_list):
    actives_table = []
    events = []
    y = -float("inf")
    actives = []
    for pt, index in edge_list:
        if y != pt.imag:
            actives_table.append(list(actives))
            events.append(y)
        if index >= 0:
            actives.append(index)
        else:
            actives.remove(~index)
        y = pt.imag
    return actives_table, events

def point_in_polygon(polygon, point):
    def x_intercept(e, y):
        pt0 = polygon[e-1]
        pt1 = polygon[e]
        if pt1.real - pt0.real == 0:
            return pt0.real
        m = (pt1.imag - pt0.imag) / (pt1.real - pt0.real)
        b = pt0.imag - (m * pt0.real)
        return (y - b) / m

    edge_list = build_edge_list(polygon)
    actives_table, events = build_scanbeam(edge_list)
    try:
        if len(point):
            return [point_in_scantable(actives_table, events, x_intercept, p) for p in point]
    except TypeError:
        return point_in_scantable(actives_table, events, x_intercept, point)

def point_in_scantable(actives_table, events, xi, point):
    beam = bisect(events, point.imag) - 1  # Binary search in sorted array.
    actives_at_y = actives_table[beam]
    total = sum([point.real > xi(e, point.imag) for e in actives_at_y])
    return bool(total % 2)

如果忽略，则扫描表的构建时间为O(n*log(n))。我们实际上是在O(log(n) + k)时间内查到的。其中n是多边形中段数的大小，k是该多边形中典型的活动边数。其他的光线追踪方法实际上需要O(n)时间。每次我们检查一个点，它迭代整个多边形。所以即使有这个明显的次优实现，它也轻而易举地打败了其他所有的。

There's a few performance tricks that could be done, for example, we can lower the time complexity to O(log(n) + log(k)) time. To do this we would implement Bentley-Ottmann into the sweep line, and rather than processing the intersections as different events, we split the lines at the intersections. We then also sort the active edges by their x-intercepts. We then know that no intersections occur during a scanbeam and since we sorted our segments (taking care to order them correctly within the scanbeam even if they start at the same initial point (you need to look at the slopes, or just compare midpoints of the segments). We then have a sorted intersection-less actives lists scanbeam table which means we can binary search into active edge list as well. Though that sounds like a lot of work for a value of k which is going to be typically 2 or maybe 4.

此外，由于这基本上变成了一个查找表和一些x截距的最小计算，它更能用GPU完成。你不再需要在多边形上循环了。所以你可以用numpy这样的东西来大量计算这些点，这样你就可以一次做所有的计算，从而提高性能。

在C语言的多边形测试中，有一个点没有使用光线投射。它可以用于重叠区域(自我交叉)，请参阅use_holes参数。

/* math lib (defined below) */
static float dot_v2v2(const float a[2], const float b[2]);
static float angle_signed_v2v2(const float v1[2], const float v2[2]);
static void copy_v2_v2(float r[2], const float a[2]);

/* intersection function */
bool isect_point_poly_v2(const float pt[2], const float verts[][2], const unsigned int nr,
                         const bool use_holes)
{
    /* we do the angle rule, define that all added angles should be about zero or (2 * PI) */
    float angletot = 0.0;
    float fp1[2], fp2[2];
    unsigned int i;
    const float *p1, *p2;

    p1 = verts[nr - 1];

    /* first vector */
    fp1[0] = p1[0] - pt[0];
    fp1[1] = p1[1] - pt[1];

    for (i = 0; i < nr; i++) {
        p2 = verts[i];

        /* second vector */
        fp2[0] = p2[0] - pt[0];
        fp2[1] = p2[1] - pt[1];

        /* dot and angle and cross */
        angletot += angle_signed_v2v2(fp1, fp2);

        /* circulate */
        copy_v2_v2(fp1, fp2);
        p1 = p2;
    }

    angletot = fabsf(angletot);
    if (use_holes) {
        const float nested = floorf((angletot / (float)(M_PI * 2.0)) + 0.00001f);
        angletot -= nested * (float)(M_PI * 2.0);
        return (angletot > 4.0f) != ((int)nested % 2);
    }
    else {
        return (angletot > 4.0f);
    }
}

/* math lib */

static float dot_v2v2(const float a[2], const float b[2])
{
    return a[0] * b[0] + a[1] * b[1];
}

static float angle_signed_v2v2(const float v1[2], const float v2[2])
{
    const float perp_dot = (v1[1] * v2[0]) - (v1[0] * v2[1]);
    return atan2f(perp_dot, dot_v2v2(v1, v2));
}

static void copy_v2_v2(float r[2], const float a[2])
{
    r[0] = a[0];
    r[1] = a[1];
}

注意:这是一个不太理想的方法，因为它包含很多对atan2f的调用，但它可能会引起阅读这个线程的开发人员的兴趣(在我的测试中，它比使用线交方法慢23倍)。

我认为下面这段代码是最好的解决方案(从这里开始):

int pnpoly(int nvert, float *vertx, float *verty, float testx, float testy)
{
  int i, j, c = 0;
  for (i = 0, j = nvert-1; i < nvert; j = i++) {
    if ( ((verty[i]>testy) != (verty[j]>testy)) &&
     (testx < (vertx[j]-vertx[i]) * (testy-verty[i]) / (verty[j]-verty[i]) + vertx[i]) )
       c = !c;
  }
  return c;
}

参数

nvert:多边形中的顶点数。是否在末端重复第一个顶点在上面的文章中已经讨论过了。 vertx, verty:包含多边形顶点的x坐标和y坐标的数组。 testx, testy:测试点的X坐标和y坐标。

它既简短又高效，适用于凸多边形和凹多边形。如前所述，您应该首先检查边界矩形，并单独处理多边形孔。

这背后的想法很简单。作者描述如下:

我从测试点水平运行一条半无限射线(增加x，固定y)，并计算它穿过多少条边。在每个十字路口，光线在内部和外部之间切换。这叫做乔丹曲线定理。

当水平射线穿过任意一条边时，变量c从0变为1，从1变为0。基本上它记录了交叉边的数量是偶数还是奇数。0表示偶数，1表示奇数。