由于这是关于这个话题最受欢迎的讨论，我将在这里补充我从2019年底到2020年初的经验。为了补充现有的答案-我的重点是找到一个准确和快速(即向量化)的解决方案。

让我们从这里最常用的答案——哈弗辛方法开始。向量化是很简单的，参见下面python中的例子:

def haversine(lat1, lon1, lat2, lon2):
    """
    Calculate the great circle distance between two points
    on the earth (specified in decimal degrees)

    All args must be of equal length.
    Distances are in meters.
    
    Ref:
    https://stackoverflow.com/questions/29545704/fast-haversine-approximation-python-pandas
    https://ipython.readthedocs.io/en/stable/interactive/magics.html
    """
    Radius = 6.371e6
    lon1, lat1, lon2, lat2 = map(np.radians, [lon1, lat1, lon2, lat2])

    dlon = lon2 - lon1
    dlat = lat2 - lat1

    a = np.sin(dlat/2.0)**2 + np.cos(lat1) * np.cos(lat2) * np.sin(dlon/2.0)**2

    c = 2 * np.arcsin(np.sqrt(a))
    s12 = Radius * c
    
    # initial azimuth in degrees
    y = np.sin(lon2-lon1) * np.cos(lat2)
    x = np.cos(lat1)*np.sin(lat2) - np.sin(lat1)*np.cos(lat2)*np.cos(dlon)
    azi1 = np.arctan2(y, x)*180./math.pi

    return {'s12':s12, 'azi1': azi1}

就精确度而言，它是最不准确的。维基百科在没有任何来源的情况下表示相对偏差平均为0.5%。我的实验显示偏差较小。以下是10万个随机点与我的库的比较，应该精确到毫米级:

np.random.seed(42)
lats1 = np.random.uniform(-90,90,100000)
lons1 = np.random.uniform(-180,180,100000)
lats2 = np.random.uniform(-90,90,100000)
lons2 = np.random.uniform(-180,180,100000)
r1 = inverse(lats1, lons1, lats2, lons2)
r2 = haversine(lats1, lons1, lats2, lons2)
print("Max absolute error: {:4.2f}m".format(np.max(r1['s12']-r2['s12'])))
print("Mean absolute error: {:4.2f}m".format(np.mean(r1['s12']-r2['s12'])))
print("Max relative error: {:4.2f}%".format(np.max((r2['s12']/r1['s12']-1)*100)))
print("Mean relative error: {:4.2f}%".format(np.mean((r2['s12']/r1['s12']-1)*100)))

输出:

Max absolute error: 26671.47m
Mean absolute error: -2499.84m
Max relative error: 0.55%
Mean relative error: -0.02%

因此，在10万对随机坐标上，平均偏差为2.5km，这可能对大多数情况都是好的。

下一个选择是Vincenty公式，精确到毫米，这取决于收敛标准，也可以向量化。它确实有在对跖点附近收敛的问题。你可以通过放宽收敛标准使其收敛于这些点，但准确度会下降到0.25%甚至更多。在对映点之外，Vincenty将提供与地理库相近的结果，相对误差小于1。平均是E-6。

这里提到的Geographiclib实际上是当前的黄金标准。它有几个实现，而且相当快，特别是如果你使用的是c++版本。

Now, if you are planning to use Python for anything above 10k points I'd suggest to consider my vectorized implementation. I created a geovectorslib library with vectorized Vincenty routine for my own needs, which uses Geographiclib as fallback for near antipodal points. Below is the comparison vs Geographiclib for 100k points. As you can see it provides up to 20x improvement for inverse and 100x for direct methods for 100k points and the gap will grow with number of points. Accuracy-wise it will be within 1.e-5 rtol of Georgraphiclib.

Direct method for 100,000 points
94.9 ms ± 25 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
9.79 s ± 1.4 s per loop (mean ± std. dev. of 7 runs, 1 loop each)

Inverse method for 100,000 points
1.5 s ± 504 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
24.2 s ± 3.91 s per loop (mean ± std. dev. of 7 runs, 1 loop each)

在我的项目中，我需要计算很多点之间的距离，所以我继续尝试优化我在这里找到的代码。平均而言，在不同的浏览器中，我的新实现的运行速度比获得最多好评的答案快2倍。

function distance(lat1, lon1, lat2, lon2) {
  var p = 0.017453292519943295;    // Math.PI / 180
  var c = Math.cos;
  var a = 0.5 - c((lat2 - lat1) * p)/2 + 
          c(lat1 * p) * c(lat2 * p) * 
          (1 - c((lon2 - lon1) * p))/2;

  return 12742 * Math.asin(Math.sqrt(a)); // 2 * R; R = 6371 km
}

您可以在这里使用我的jsPerf并查看结果。

最近我需要在python中做同样的事情，所以这里是一个python实现:

from math import cos, asin, sqrt, pi

def distance(lat1, lon1, lat2, lon2):
    p = pi/180
    a = 0.5 - cos((lat2-lat1)*p)/2 + cos(lat1*p) * cos(lat2*p) * (1-cos((lon2-lon1)*p))/2
    return 12742 * asin(sqrt(a)) #2*R*asin...

为了完整起见:维基百科上的Haversine。

可能有一个更简单、更正确的解决方案:地球的周长在赤道上是40000公里，在格林威治(或任何经度)周期上约为37000公里。因此:

pythagoras = function (lat1, lon1, lat2, lon2) {
   function sqr(x) {return x * x;}
   function cosDeg(x) {return Math.cos(x * Math.PI / 180.0);}

   var earthCyclePerimeter = 40000000.0 * cosDeg((lat1 + lat2) / 2.0);
   var dx = (lon1 - lon2) * earthCyclePerimeter / 360.0;
   var dy = 37000000.0 * (lat1 - lat2) / 360.0;

   return Math.sqrt(sqr(dx) + sqr(dy));
};

我同意它应该被微调，我自己说过它是一个椭球，所以半径乘以余弦值是不同的。但它更准确一点。与谷歌map相比，误差明显减小。

我通过简化公式来简化计算。

下面是Ruby版本:

include Math
earth_radius_mi = 3959
radians = lambda { |deg| deg * PI / 180 }
coord_radians = lambda { |c| { :lat => radians[c[:lat]], :lng => radians[c[:lng]] } }

# from/to = { :lat => (latitude_in_degrees), :lng => (longitude_in_degrees) }
def haversine_distance(from, to)
  from, to = coord_radians[from], coord_radians[to]
  cosines_product = cos(to[:lat]) * cos(from[:lat]) * cos(from[:lng] - to[:lng])
  sines_product = sin(to[:lat]) * sin(from[:lat])
  return earth_radius_mi * acos(cosines_product + sines_product)
end

你也可以使用像geolib这样的模块:

安装方法:

$ npm install geolib

使用方法:

import { getDistance } from 'geolib'

const distance = getDistance(
    { latitude: 51.5103, longitude: 7.49347 },
    { latitude: "51° 31' N", longitude: "7° 28' E" }
)

console.log(distance)

文档: https://www.npmjs.com/package/geolib

function getDistanceFromLatLonInKm(position1, position2) {
    "use strict";
    var deg2rad = function (deg) { return deg * (Math.PI / 180); },
        R = 6371,
        dLat = deg2rad(position2.lat - position1.lat),
        dLng = deg2rad(position2.lng - position1.lng),
        a = Math.sin(dLat / 2) * Math.sin(dLat / 2)
            + Math.cos(deg2rad(position1.lat))
            * Math.cos(deg2rad(position2.lat))
            * Math.sin(dLng / 2) * Math.sin(dLng / 2),
        c = 2 * Math.atan2(Math.sqrt(a), Math.sqrt(1 - a));
    return R * c;
}

console.log(getDistanceFromLatLonInKm(
    {lat: 48.7931459, lng: 1.9483572},
    {lat: 48.827167, lng: 2.2459745}
));