21
votes

I need to calculate the shortest distance from a lat/lng GPS point P to a line segment described by 2 other lat/lng GPS points A and B.

'Cross-track distance' helps me to calculate the shortest distance between P and the great circle described by A and B.

However, this is not what I want. I need need the distance between P and the line segment of A-B, not the entire great circle.

I have used the following implementation from http://www.movable-type.co.uk/scripts/latlong.html

Formula:    dxt = asin( sin(δ13) ⋅ sin(θ13−θ12) ) ⋅ R
where:
δ13 is (angular) distance from start point to third point
θ13 is (initial) bearing from start point to third point
θ12 is (initial) bearing from start point to end point
R is the earth’s radius

The following images hopefully demonstrate the problem I am trying to solve: Cross-Track distance Correct Cross-Track distance Incorrect

In the first image the Cross-Track distance, indicated by the green line is correct and indeed the shortest distance to the line segment AB.

In the second image the problem with cross-track distance is shown, In this case I would want the shortest distance to be the simple distance AP, but Cross-Track distance gives me the distance indicated by the red line.

How do I change my algoritm to take this into account, or check whether or not point X is within AB. Is it possible to do this computationally? Or is iterative the only possible (expensive) solution? (take N points along AB and calculate the min distance from P to all these points)

For simplicity purposes all lines in the images are straight. In reality, these are minor arcs on a great circle

7
What are the maximum distances in that application? some hundred meters, or up to some thousand kilomters?AlexWien
The line segments are waypoints in a (driven) route. I would say the distances between them would be in the 100-1000m range, roughly.ChrisDekker
Did you already translate the cross track distance to matlab? If so could you put the code into your question?Daniel
In the plane, you'd just check if all angles in the ABP triangle were <= 90 degrees. Is there something similar that could be applied from spherical trig?Chad Kennedy
Sure Chad, was thinking the same thing at first, but I could come up with several scenario's where that would not produce accurate results, including points spanning different hemispheres and instances where the minor arc segment was very smallChrisDekker

7 Answers

27
votes

First, some nomenclature:
Our arc is drawn from p1 to p2.
Our third point is p3.
The imaginary point that intersects the great circle is p4.
p1 is defined by lat1,lon1; p2 by lat2,lon2; etc.
dis12 is the distance from p1 to p2; etc.
bear12 is the bearing from p1 to p2; etc.
dxt is cross-track distance.
dxa is cross-arc distance, our goal!

Notice that the cross-track formula relies on the relative bearing, bear13-bear12

We have 3 cases to deal with.

Case 1: The relative bearing is obtuse. So, dxa=dis13.

Case 1

Case 2.1: The relative bearing is acute, AND p4 falls on our arc. So, dxa=dxt.

Case 2.1

Case 2.2: The relative bearing is acute,AND p4 falls beyond our arc. So, dxa=dis23

enter image description here

The algorithm:

Step 1: If relative bearing is obtuse, dxa=dis13
Done!
Step 2: If relative bearing is acute:
2.1: Find dxt.
2.3: Find dis12.
2.4: Find dis14.
2.4: If dis14>dis12, dxa=dis23.
Done!
2.5: If we reach here, dxa=abs(dxt)

MATLAB code:

function [ dxa ] = crossarc( lat1,lon1,lat2,lon2,lat3,lon3 )
%// CROSSARC Calculates the shortest distance in meters 
%// between an arc (defined by p1 and p2) and a third point, p3.
%// Input lat1,lon1,lat2,lon2,lat3,lon3 in degrees.
    lat1=deg2rad(lat1); lat2=deg2rad(lat2); lat3=deg2rad(lat3);
    lon1=deg2rad(lon1); lon2=deg2rad(lon2); lon3=deg2rad(lon3);

    R=6371000; %// Earth's radius in meters
    %// Prerequisites for the formulas
    bear12 = bear(lat1,lon1,lat2,lon2);
    bear13 = bear(lat1,lon1,lat3,lon3);
    dis13 = dis(lat1,lon1,lat3,lon3);

    diff = abs(bear13-bear12);
    if diff > pi
        diff = 2 * pi - diff;
    end
    %// Is relative bearing obtuse?
    if diff>(pi/2)
        dxa=dis13;
    else
        %// Find the cross-track distance.
        dxt = asin( sin(dis13/R)* sin(bear13 - bear12) ) * R;

        %// Is p4 beyond the arc?
        dis12 = dis(lat1,lon1,lat2,lon2);
        dis14 = acos( cos(dis13/R) / cos(dxt/R) ) * R;
        if dis14>dis12
            dxa=dis(lat2,lon2,lat3,lon3);
        else
            dxa=abs(dxt);
        end   
    end
end

function [ d ] = dis( latA, lonA, latB, lonB )
%DIS Finds the distance between two lat/lon points.
R=6371000;
d = acos( sin(latA)*sin(latB) + cos(latA)*cos(latB)*cos(lonB-lonA) ) * R;
end

function [ b ] = bear( latA,lonA,latB,lonB )
%BEAR Finds the bearing from one lat/lon point to another.
b=atan2( sin(lonB-lonA)*cos(latB) , ...
    cos(latA)*sin(latB) - sin(latA)*cos(latB)*cos(lonB-lonA) );
end

Sample outputs: Demonstrate all cases. See maps below.

>> crossarc(-10.1,-55.5,-15.2,-45.1,-10.5,-62.5)
ans =
   7.6709e+05
>> crossarc(40.5,60.5,50.5,80.5,51,69)
ans =
   4.7961e+05
>> crossarc(21.72,35.61,23.65,40.7,25,42)
ans =
   1.9971e+05

Those same outputs on the map!:

Demonstrates case 1:

Case 1 on map

Demonstrates case 2.1:

Case 2.1 on map

Demonstrates case 2.2:

Case 2.2 on map

Credit to: http://www.movable-type.co.uk/scripts/latlong.html
for the formulas
and: http://www.darrinward.com/lat-long/?id=1788764
for generating the map images.

2
votes

Adding a Java version to wdickerson answer:

public static double pointToLineDistance(double lon1, double lat1, double lon2, double lat2, double lon3, double lat3) {
    lat1 = Math.toRadians(lat1);
    lat2 = Math.toRadians(lat2);
    lat3 = Math.toRadians(lat3);
    lon1 = Math.toRadians(lon1);
    lon2 = Math.toRadians(lon2);
    lon3 = Math.toRadians(lon3);

    // Earth's radius in meters
    double R = 6371000;

    // Prerequisites for the formulas
    double bear12 = bear(lat1, lon1, lat2, lon2);
    double bear13 = bear(lat1, lon1, lat3, lon3);
    double dis13 = dis(lat1, lon1, lat3, lon3);

    // Is relative bearing obtuse?
    if (Math.abs(bear13 - bear12) > (Math.PI / 2))
        return dis13;

    // Find the cross-track distance.
    double dxt = Math.asin(Math.sin(dis13 / R) * Math.sin(bear13 - bear12)) * R;

    // Is p4 beyond the arc?
    double dis12 = dis(lat1, lon1, lat2, lon2);
    double dis14 = Math.acos(Math.cos(dis13 / R) / Math.cos(dxt / R)) * R;
    if (dis14 > dis12)
        return dis(lat2, lon2, lat3, lon3);
    return Math.abs(dxt);
}

private static double dis(double latA, double lonA, double latB, double lonB) {
    double R = 6371000;
    return Math.acos(Math.sin(latA) * Math.sin(latB) + Math.cos(latA) * Math.cos(latB) * Math.cos(lonB - lonA)) * R;
}

private static double bear(double latA, double lonA, double latB, double lonB) {
    // BEAR Finds the bearing from one lat / lon point to another.
    return Math.atan2(Math.sin(lonB - lonA) * Math.cos(latB), Math.cos(latA) * Math.sin(latB) - Math.sin(latA) * Math.cos(latB) * Math.cos(lonB - lonA));
}
2
votes

And adding a python translation of Sga's implementation:

    def bear(latA, lonA, latB, lonB):
        # BEAR Finds the bearing from one lat / lon point to another.
        return math.atan2(
            math.sin(lonB - lonA) * math.cos(latB),
            math.cos(latA) * math.sin(latB) - math.sin(latA) * math.cos(latB) * math.cos(lonB - lonA)
        )


    def pointToLineDistance(lon1, lat1, lon2, lat2, lon3, lat3):
        lat1 = math.radians(lat1)
        lat2 = math.radians(lat2)
        lat3 = math.radians(lat3)
        lon1 = math.radians(lon1)
        lon2 = math.radians(lon2)
        lon3 = math.radians(lon3)
        R = 6378137

        bear12 = bear(lat1, lon1, lat2, lon2)
        bear13 = bear(lat1, lon1, lat3, lon3)
        dis13 = distance( (lat1, lon1), (lat3, lon3)).meters

        # Is relative bearing obtuse?
        if math.fabs(bear13 - bear12) > (math.pi / 2):
            return dis13

        # Find the cross-track distance.
        dxt = math.asin(math.sin(dis13 / R) * math.sin(bear13 - bear12)) * R

        # Is p4 beyond the arc?
        dis12 = distance((lat1, lon1), (lat2, lon2)).meters
        dis14 = math.acos(math.cos(dis13 / R) / math.cos(dxt / R)) * R
        if dis14 > dis12:
            return distance((lat2, lon2), (lat3, lon3)).meters
        return math.fabs(dxt)
1
votes

For 100 - 1000m spherical problems, it is easy to just convert to cartesian space, using a equirectangular projection.
Then it continues with school mathematics:
Use the function "distance from line segment" which is easy to find ready implemented. This fucntion uses (and sometimes returns) a relative forward/backward position for the projected point X on the line A,B. The value is

  • in the interval [0,1] if the projected point is inside the line segment.
  • it is negative if X is outside before A,
  • it is >1 if outside after B.
    If the relative position is between 0,1 the normal distance is taken, if outside the shorter distance of the both start and line-end points, A,B.

An example of such / or very similar an cartesian implementaion is Shortest distance between a point and a line segment

0
votes
/**
 * Calculates the euclidean distance from a point to a line segment.
 *
 * @param v     the point
 * @param a     start of line segment
 * @param b     end of line segment 
 * @return      an array of 2 doubles:
 *              [0] distance from v to the closest point of line segment [a,b],
 *              [1] segment coeficient of the closest point of the segment.
 *              Coeficient values < 0 mean the closest point is a.
 *              Coeficient values > 1 mean the closest point is b.
 *              Coeficient values between 0 and 1 mean how far along the segment the closest point is.
 *
 * @author         Afonso Santos
 */
public static
double[]
distanceToSegment( final R3 v, final R3 a, final R3 b )
{
    double[] results    = new double[2] ;

    final R3     ab_    = b.sub( a ) ;
    final double ab     = ab_.modulus( ) ;

    final R3     av_    = v.sub( a ) ;
    final double av     = av_.modulus( ) ;

    if (ab == 0.0)                       // a and b coincide
    {
        results[0] = av ;                // Distance
        results[1] = 0.0 ;               // Segment coeficient.
    }
    else
    {
        final double avScaProjAb  = av_.dot(ab_) / ab ;
        final double abCoeficient = results[1] = avScaProjAb / ab ;

        if (abCoeficient <= 0.0)                 // Point is before start of the segment ?
            results[0] = av ;                    // Use distance to start of segment.
        else if (abCoeficient >= 1.0)            // Point is past the end of the segment ?
            results[0] = v.sub( b ).modulus() ;    // Use distance to end of segment.
        else                                       // Point is within the segment's start/end perpendicular boundaries.
        {
            if (avScaProjAb >= av)                    // Test to avoid machine float representation epsilon rounding errors that would result in expection on sqrt.
                results[0] = 0.0 ;                    // a, b and v are colinear.
            else
                results[0] = Math.sqrt( av * av - avScaProjAb * avScaProjAb ) ;        // Perpendicular distance from point to segment.
        }
    }

    return results ;
}

the above method requires cartesian 3D space arguments and you asked to use lat/lon arguments. To do the conversion use

/**
 * Calculate 3D vector (from center of earth).
 * 
 * @param latDeg    latitude (degrees)
 * @param lonDeg    longitude (degrees)
 * @param eleMtr    elevation (meters)
 * @return          3D cartesian vector (from center of earth).
 * 
 * @author          Afonso Santos
 */
public static
R3
cartesian( final double latDeg, final double lonDeg, final double eleMtr )
{
    return versor( latDeg, lonDeg ).scalar( EARTHMEANRADIUS_MTR + eleMtr ) ;
}

For the rest of the 3D/R3 code or how to calculate distance to a path/route/track check https://sourceforge.net/projects/geokarambola/

0
votes

Adding an ObjectiveC translation of wdickerson implementation:

#define DEGREES_RADIANS(angle) ((angle) / 180.0 * M_PI)
#define RADIANS_DEGREES(angle) ((angle) / M_PI * 180)

- (double)crossArcFromCoord:(CLLocationCoordinate2D)fromCoord usingArcFromCoord:(CLLocationCoordinate2D)arcCoord1 toArcCoord:(CLLocationCoordinate2D)arcCoord2 {

        fromCoord.latitude = DEGREES_RADIANS(fromCoord.latitude);
        fromCoord.longitude = DEGREES_RADIANS(fromCoord.longitude);

        arcCoord1.latitude = DEGREES_RADIANS(arcCoord1.latitude);
        arcCoord1.longitude = DEGREES_RADIANS(arcCoord1.longitude);

        arcCoord2.latitude = DEGREES_RADIANS(arcCoord2.latitude);
        arcCoord2.longitude = DEGREES_RADIANS(arcCoord2.longitude);

        double R = 6371000; // Earth's radius in meters

        // Prerequisites for the formulas
        double bear12 = [self bearFromCoord:arcCoord1 toCoord:arcCoord2];
        double bear13 = [self bearFromCoord:arcCoord1 toCoord:fromCoord];

        double dis13 = [self distFromCoord:arcCoord1 toCoord:fromCoord];

        double diff = fabs(bear13 - bear12);

        if (diff > M_PI) {
            diff = 2 * M_PI - diff;
        }

        // Is relative bearing obtuse?
        if (diff > (M_PI/2)) {
            return dis13;
        }

        // Find the cross-track distance
        double dxt = asin(sin(dis13 / R) * sin(bear13 - bear12)) * R;

        // Is p4 beyond the arc?
        double dis12 = [self distFromCoord:arcCoord1 toCoord:arcCoord2];
        double dis14 = acos(cos(dis13 / R) / cos(dxt / R)) * R;

        if (dis14 > dis12) {
            return [self distFromCoord:arcCoord2 toCoord:fromCoord];
        }

        return fabs(dxt);
    }

    - (double)distFromCoord:(CLLocationCoordinate2D)coord1 toCoord:(CLLocationCoordinate2D)coord2 {

        double R = 6371000;

        return acos(sin(coord1.latitude) * sin(coord2.latitude) + cos(coord1.latitude) * cos(coord2.latitude) * cos(coord2.longitude - coord2.longitude)) * R;
    }

    - (double)bearFromCoord:(CLLocationCoordinate2D)fromCoord toCoord:(CLLocationCoordinate2D)toCoord {

        return atan2(sin(toCoord.longitude - fromCoord.longitude) * cos(toCoord.latitude),
                     cos(fromCoord.latitude) * sin(toCoord.latitude) - (sin(fromCoord.latitude) * cos(toCoord.latitude) * cos(toCoord.longitude - fromCoord.longitude)));
     }
0
votes

Adding a python+numpy implementation (now you can pass your longitudes and latitudes as arrays and compute all your distances simultaneously without loops).

def _angularSeparation(long1, lat1, long2, lat2):
    """All radians
    """
    t1 = np.sin(lat2/2.0 - lat1/2.0)**2
    t2 = np.cos(lat1)*np.cos(lat2)*np.sin(long2/2.0 - long1/2.0)**2
    _sum = t1 + t2

    if np.size(_sum) == 1:
        if _sum < 0.0:
            _sum = 0.0
    else:
        _sum = np.where(_sum < 0.0, 0.0, _sum)

    return 2.0*np.arcsin(np.sqrt(_sum))


def bear(latA, lonA, latB, lonB):
    """All radians
    """
    # BEAR Finds the bearing from one lat / lon point to another.
    result = np.arctan2(np.sin(lonB - lonA) * np.cos(latB),
                        np.cos(latA) * np.sin(latB) - np.sin(latA) * np.cos(latB) * np.cos(lonB - lonA)
                        )

    return result


def pointToLineDistance(lon1, lat1, lon2, lat2, lon3, lat3):
    """All radians
    points 1 and 2 define an arc segment,
    this finds the distance of point 3 to the arc segment. 
    """

    result = lon1*0
    needed = np.ones(result.size, dtype=bool)

    bear12 = bear(lat1, lon1, lat2, lon2)
    bear13 = bear(lat1, lon1, lat3, lon3)
    dis13 = _angularSeparation(lon1, lat1, lon3, lat3)

    # Is relative bearing obtuse?
    diff = np.abs(bear13 - bear12)
    if np.size(diff) == 1:
        if diff > np.pi:
            diff = 2*np.pi - diff
        if diff > (np.pi / 2):
            return dis13
    else:
        solved = np.where(diff > (np.pi / 2))[0]
        result[solved] = dis13[solved]
        needed[solved] = 0
    
    # Find the cross-track distance.
    dxt = np.arcsin(np.sin(dis13) * np.sin(bear13 - bear12))

    # Is p4 beyond the arc?
    dis12 = _angularSeparation(lon1, lat1, lon2, lat2)
    dis14 = np.arccos(np.cos(dis13) / np.cos(dxt))
    if np.size(dis14) == 1:
        if dis14 > dis12:
            return _angularSeparation(lon2, lat2, lon3, lat3)
    else:
        solved = np.where(dis14 > dis12)[0]
        result[solved] = _angularSeparation(lon2[solved], lat2[solved], lon3[solved], lat3[solved])

    if np.size(lon1) == 1:
        return np.abs(dxt)
    else:
        result[needed] = np.abs(dxt[needed])
        return result