imu position without GPS or camera

Question

I have a IMU that has 3-axis accelerator, 3-axis magnetometer, 3-axis gyroscope and row, yaw, pitch value. I want to get the location of the IMU coordinate(the beginning point is (0,0,0)) but I know just using double integration will has dead reckoning problem. And I found a lot of paper just talking about combining IMU with GPS or camera by using Kalman filter. Is it possible that I just use a IMU to get a slightly precise position data? Because in the future work I will use multiple IMUs bounded on human arms to increase the accuracy.

Answer 1

No, it is not possible to eliminate the cumulative position error caused by sensor noise and bias without using an additional sensor which can report any kind of position measurement.

Even the best sensors and filtering will not be able to eliminate in a closed-loop fashion the position error.

Answer 2

Of course, it is possible. You need to go for a sensor fusion algorithm, that could be Kalman Filter or Complementary Filter.

I personally found very useful the procedure described at the following link, in order to have a total 3D estimation, without suffering any gimbal lock or other problems.

EDIT: I am reading only now that you're asking for position estimation, and not orientation estimation. My answer refers to the latter. For a precise position estimation unfortunately you would need other kind of sensors (if your robot has wheels (not clear from the question), you could use encoders instead of GPS).

Answer 3

The answer to this question greatly depends on your acceptable error, and budget.

As the other two answers have stated, it is practically impossible to dead-reckon the position without directly observing the position with another sensor; this is however based on assumptions about your acceptable error and IMU selection.

If your acceptable error is large enough, and the time period of which you need operation is small enough, you could possibly estimate the bias of the MEMS sensors as an average measurements during a "calibration" period.

e.g. Don't move the IMU for 5 seconds, record the average accelerations and angular rates, and use that as the bias over the operating period of the next 5 seconds. Do the double-integration for the accelerometer and angular integration (I'd suggest Quaternion integration) for the gyro, but subtract each sensors calculated bias. I must stress that this will only work for a few seconds for most cheap MEMS IMUs.

Alternatively, if you use an incredibly expensive IMU (e.g. naval IMU, up to $2m, 35 kg) you could probably dead-reckon to within acceptable errors for a substantial time period.