This is a work in progress
I have been working in personal navigation since 1994. In that time I have had to manage the calibration of commercial grade sensors (accelerometers, magnetometers, barometric altimeters, and gyroscopes).
Magnetic compasses are used widely in personal navigation devices. We developed a number of digital compasses for use in those devices, and also as stand-alone devices. In 2004 our precision compass averaged better than 0.2 degrees sigma when calibrated in the factory. In 2006 I developed a field calibration method for our navigation devices that did not require a calibration table. Users could attain near-factory accuracy in the field .
Most of my embedded system programming has been done in C. However, C syntax for managing Matrix algebra is quite awkward. The sample code here would get bloated from all the auxiliary functions needed to operate on matrices and vectors. Python (NumPy, SciPy) has those already. To compensate, I have included test vectors for input and sample output so that a port to C, or MATLAB, or your pet language can be validated. If you do not yet have Python try Python(X,Y) version 188.8.131.52 (latest as of Feb 2017). All the Python code here was written with it. It has the IDE (Spyder), plotting (MatPlotLib), and NumPy all in one consistent package.
The earth's magnetic field should be locally constant. Whatever the orientation of a platform with magnetic sensors, the reported magnitude of the magnetic field should be the same. In two dimensions, an unperturbed compass will report a constant field in all directions. Plotting the data will produce a circle centered at the origin. But when those sensors are stuffed into a device that has nearby ferromagnetic materials, the local magnetic field may be perturbed so that we do not see that symmetry in the response. Local material may be magnetized. This can add a constant vector to the earth's field. It will enhance the field when the earth's field is aligned with the local field, but will subtract from that field when it is aligned against the local field. That set of data will still be circular, but shifted away from the center by the magnitude of the perturbing field. This is a Hard Iron error.
Even un-magnetized materials have an effect. They serve as conduits for magnetic lines of flux. The resulting field can be enhanced or shielded along one direction. In two dimensions, the plotted data will be elliptical. This is a Soft Iron error. The major axes will not necessarily be aligned with our coordinate system.
It is quite common to have significant errors from both hard and soft perturbations. In two dimensions that would be an off-center tilted ellipse. In 3D it is an ellipsoid. The sample diagram of a 3D ellipsoid is taken from the 1985 patent of Marchent and Foster . They developed the classic process of calibrating a 3D compass as follows:
There is no constraint that forces the above equation converge on an ellipsoid. With sparse or noisy data we could find ourselves with a 3D hyperbola. Refinements can force the solution to be an ellipsoid.. These methods can pull a solution out of data where other algorithms fail. Nevertheless, bad, incomplete, or noisy data will always degrade the results. These techniques are moot when the goal is a precision compass.
Notice that there is no mention of a reference platform or calibration table in the above process. That is because you do not need a reference platform to calibrate a magnetic compass. A calibration table is needed in the factory to develop algorithms and to validate their effectiveness. But in the field, users may apply these powerful calibration techniques without the expense or trouble of using one.
A well calibrated compass has near spherical symmetry. After calibration in one coordinate system, the user should be able to achieve the same accuracy in another. For example, calibrate the system with the z axis up and y axis forward. Next redefine the vertical as the x axis with -z forward. Then measure the accuracy without recalibration. The compass should remain within specifications. We required such a test as a part of our compass validation process.
Hard Iron Error from Local Permanent Magnet
Soft Iron Error from Local Ferromagnetic Materials
3D Ellipsoid from Marchent & Foster patent, 1985
I worked in analysis of torpedo sonar data at the Point Loma Naval Oceans Systems Center (1982-84). I also was part of a startup group at RD Instruments (1984-89) that developed the first doppler sonar ocean current measurment instruments (ADCP). There are generally more than three sonar beams in these systems and they definitely are not aligned with the instrument's xyz coordinates. In addition, the instrument body itself is not aligned with the earth's coordinates (shipboard systems and bouys). I developed a least squares method to project an arbritrary array of instrument sensors onto a given definition of x, y, and z. So, when I encountered the 3D magnetic sensor array I immediately recognized that there was no theoretical constraint on the mounting of the sensors. A valid body-centered coordinate system could always be projected out from the sensor coordinates. In addition, once that coordinate system was defined it would be possible to project a north pointing vector onto a horizontal plane in that coordinate system.
We get our direction by projecting the forward vector of our instrument onto the horizontal plane and measuring the angle between the projection and the north pointing vector. At high pitch angles the projection gets noisy; at 90 degrees it is undefined; at 91 degrees it flips sign. We handle these extreme cases by redefining what we consider forward. For example, patent 7103471 handles the case of a person-mounted compass when the user lies down on his stomach. In our person-mounted navigation systems we modeled the sensor as in the center of a cube with the sensor coordinate axes centered on the six faces. The user could specify any of six faces as the down direction. Of the four vertical faces, he could choose one for forward. Twenty four unique mounting orientations seemed to be enough for any user.
I developed our iterative compass calibration algorithm from 1994-1998 at Point Research, Inc. I saw that a polynomial solution took a lot of work. I was interested in something that could be developed quickly and easily. I decided on an iterative process where the parameters are jiggled a little at each step. The best fit parameters are selected for the next round. The iterations are stopped with a convergence criterion that checks to see if there is any significant improvement.
But that is not all! I also dispensed completely with the polynomial and went directly to the offsets and transformation matrix. That is, the three offsets and nine elements of the 3x3 transformation matrix are the parameters I seek, not the ellipsoid polynomial coefficients. For the classical solutions, the matrix is symmetric, so there are only nine unique parameters. For high precision compass calibrations, we use all twelve parameters.
A side feature of this approach is that it circumvents the need to solve the eigenvalue/eigenvector problems. While this is a one-liner in MATLAB, it is not as easy in small embedded systems.
Classic calibration of a magnetic compass can give you a compass with as low as a one degree error.
This is about as good as you can get when you use only the magnetic field samples.
Bob M., a co-worker at Point Research, applied additional
constraints using a calibration table. He was able to achieve better than one degree accuracy in the factory.
Gnepf and Weilenmann  showed how to include additional constraints from the accelerometers.
Their constraint is that the angle between the magnetic field and the gravitational field be constant.
I have tried a number of constraints that mix the accelerometer and magnetometer data:
Below, a sample calculation on noisy, simulated data demonstrates the possibilities for precision iterative calibrations. Reproducable accuracy of better than a quarter degree sigma is possible.
Blue: Original Error
Orange: Hard Iron Error Corrected (3 parameters)
Yellow: Hard plus Soft Iron Error Corrected (9 parameters)
Green: Full Precision (12 parameters)
Iterative convergence is fairly slow, but quite acceptable for factory calibrations. I tried moving the calibration to the embedded system but it took minutes and minutes to converge. One final improvement by another co-worker, Eric E., sped up the calibration algorithm. He linearized the problem. Useable answers came with only a few iterations. I saw that we could now put the calibration in our compasses. The new calibrations took seconds to produce fractional degree results. Users could attain near-factory accuracy in the field .