我知道这是旧的，但这里是一个在Cocoa实现的光线投射算法，如果有人感兴趣的话。不确定这是最有效的方法，但它可能会帮助别人。

- (BOOL)shape:(NSBezierPath *)path containsPoint:(NSPoint)point
{
    NSBezierPath *currentPath = [path bezierPathByFlatteningPath];
    BOOL result;
    float aggregateX = 0; //I use these to calculate the centroid of the shape
    float aggregateY = 0;
    NSPoint firstPoint[1];
    [currentPath elementAtIndex:0 associatedPoints:firstPoint];
    float olderX = firstPoint[0].x;
    float olderY = firstPoint[0].y;
    NSPoint interPoint;
    int noOfIntersections = 0;

    for (int n = 0; n < [currentPath elementCount]; n++) {
        NSPoint points[1];
        [currentPath elementAtIndex:n associatedPoints:points];
        aggregateX += points[0].x;
        aggregateY += points[0].y;
    }

    for (int n = 0; n < [currentPath elementCount]; n++) {
        NSPoint points[1];

        [currentPath elementAtIndex:n associatedPoints:points];
        //line equations in Ax + By = C form
        float _A_FOO = (aggregateY/[currentPath elementCount]) - point.y;  
        float _B_FOO = point.x - (aggregateX/[currentPath elementCount]);
        float _C_FOO = (_A_FOO * point.x) + (_B_FOO * point.y);

        float _A_BAR = olderY - points[0].y;
        float _B_BAR = points[0].x - olderX;
        float _C_BAR = (_A_BAR * olderX) + (_B_BAR * olderY);

        float det = (_A_FOO * _B_BAR) - (_A_BAR * _B_FOO);
        if (det != 0) {
            //intersection points with the edges
            float xIntersectionPoint = ((_B_BAR * _C_FOO) - (_B_FOO * _C_BAR)) / det;
            float yIntersectionPoint = ((_A_FOO * _C_BAR) - (_A_BAR * _C_FOO)) / det;
            interPoint = NSMakePoint(xIntersectionPoint, yIntersectionPoint);
            if (olderX <= points[0].x) {
                //doesn't matter in which direction the ray goes, so I send it right-ward.
                if ((interPoint.x >= olderX && interPoint.x <= points[0].x) && (interPoint.x > point.x)) {  
                    noOfIntersections++;
                }
            } else {
                if ((interPoint.x >= points[0].x && interPoint.x <= olderX) && (interPoint.x > point.x)) {
                     noOfIntersections++;
                } 
            }
        }
        olderX = points[0].x;
        olderY = points[0].y;
    }
    if (noOfIntersections % 2 == 0) {
        result = FALSE;
    } else {
        result = TRUE;
    }
    return result;
}

bobobobo引用的Eric Haines的文章真的很棒。特别有趣的是比较算法性能的表格;角度求和法和其他方法比起来真的很差。同样有趣的是，使用查找网格将多边形进一步细分为“in”和“out”扇区的优化可以使测试非常快，即使是在> 1000条边的多边形上。

不管怎样，现在还为时过早，但我的投票倾向于“交叉”方法，我认为这几乎就是Mecki所描述的。然而，我发现大卫·伯克(David Bourke)对它进行了最简洁的描述和编纂。我喜欢它不需要真正的三角函数，它适用于凸和凹，而且随着边数的增加，它的表现也相当不错。

顺便说一下，这是Eric Haines文章中的一个性能表，在随机多边形上进行测试。

                       number of edges per polygon
                         3       4      10      100    1000
MacMartin               2.9     3.2     5.9     50.6    485
Crossings               3.1     3.4     6.8     60.0    624
Triangle Fan+edge sort  1.1     1.8     6.5     77.6    787
Triangle Fan            1.2     2.1     7.3     85.4    865
Barycentric             2.1     3.8    13.8    160.7   1665
Angle Summation        56.2    70.4   153.6   1403.8  14693

Grid (100x100)          1.5     1.5     1.6      2.1      9.8
Grid (20x20)            1.7     1.7     1.9      5.7     42.2
Bins (100)              1.8     1.9     2.7     15.1    117
Bins (20)               2.1     2.2     3.7     26.3    278

计算点p与每个多边形顶点之间的有向角和。如果总倾斜角是360度，那么这个点在里面。如果总数为0，则点在外面。

我更喜欢这种方法，因为它更健壮，对数值精度的依赖更小。

计算交集数量的均匀性的方法是有限的，因为你可以在计算交集数量的过程中“击中”一个顶点。

编辑:顺便说一下，这种方法适用于凹凸多边形。

编辑:我最近在维基百科上找到了一篇关于这个话题的完整文章。

David Segond's answer is pretty much the standard general answer, and Richard T's is the most common optimization, though therre are some others. Other strong optimizations are based on less general solutions. For example if you are going to check the same polygon with lots of points, triangulating the polygon can speed things up hugely as there are a number of very fast TIN searching algorithms. Another is if the polygon and points are on a limited plane at low resolution, say a screen display, you can paint the polygon onto a memory mapped display buffer in a given colour, and check the color of a given pixel to see if it lies in the polygons.

像许多优化一样，这些优化是基于特定情况而不是一般情况，并且基于摊销时间而不是单次使用产生效益。

在这个领域工作，我发现约瑟夫·奥鲁克斯的《计算几何》在C' ISBN 0-521-44034-3是一个很大的帮助。

下面是Rust版本的@nirg答案(Philipp Lenssen javascript版本) 我给出这个答案是因为我从这个网站得到了很多帮助，我翻译javascript版本rust作为一个练习，希望可以帮助一些人，最后一个原因是，在我的工作中，我会把这段代码翻译成一个wasm，以提高我的画布的性能，这是一个开始。我的英语很差……,请原谅我 ｀

pub struct Point {
    x: f32,
    y: f32,
}
pub fn point_is_in_poly(pt: Point, polygon: &Vec<Point>) -> bool {
    let mut is_inside = false;

    let max_x = polygon.iter().map(|pt| pt.x).reduce(f32::max).unwrap();
    let min_x = polygon.iter().map(|pt| pt.x).reduce(f32::min).unwrap();
    let max_y = polygon.iter().map(|pt| pt.y).reduce(f32::max).unwrap();
    let min_y = polygon.iter().map(|pt| pt.y).reduce(f32::min).unwrap();

    if pt.x < min_x || pt.x > max_x || pt.y < min_y || pt.y > max_y {
        return is_inside;
    }

    let len = polygon.len();
    let mut j = len - 1;

    for i in 0..len {
        let y_i_value = polygon[i].y > pt.y;
        let y_j_value = polygon[j].y > pt.y;
        let last_check = (polygon[j].x - polygon[i].x) * (pt.y - polygon[i].y)
            / (polygon[j].y - polygon[i].y)
            + polygon[i].x;
        if y_i_value != y_j_value && pt.x < last_check {
            is_inside = !is_inside;
        }
        j = i;
    }
    is_inside
}


let pt = Point {
    x: 1266.753,
    y: 97.655,
};
let polygon = vec![
    Point {
        x: 725.278,
        y: 203.586,
    },
    Point {
        x: 486.831,
        y: 441.931,
    },
    Point {
        x: 905.77,
        y: 445.241,
    },
    Point {
        x: 1026.649,
        y: 201.931,
    },
];
let pt1 = Point {
    x: 725.278,
    y: 203.586,
};
let pt2 = Point {
    x: 872.652,
    y: 321.103,
};
println!("{}", point_is_in_poly(pt, &polygon));// false
println!("{}", point_is_in_poly(pt1, &polygon)); // true
println!("{}", point_is_in_poly(pt2, &polygon));// true

`

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。