Possible division by zero in PreintegratedImuMeasurements::integrateMeasurement (ImuFactor.cpp)

In ImuFactor.cpp for the covariance updating there are divisions by dt, which can be 0.0. Now I get that in general there’d be no reason to update with a 0.0 timestep, but in my particular case this sometimes can happen..

So at line 72 and 73 in ImuFactor.cpp there is

  // (1/dt) allows to pass from continuous time noise to discrete time noise
  preintMeasCov_ = A * preintMeasCov_ * A.transpose();
  preintMeasCov_.noalias() += B * (aCov / dt) * B.transpose();
  preintMeasCov_.noalias() += C * (wCov / dt) * C.transpose();

and since updating of the tangent space vector has no problems with dt=0.0, I think it can be confusing to have an ok state, a covariance matrix with nan’s and no warning or error.

This can easily be prevented by doing

if (dt > 0.0)
{
  obj.integrateMeasurement(acc,gyr,dt);
}

But I guess we might want move this check to the point where we update the covariance matrix so you can actually call this function with a 0.0 timestep. Or just prevent updating anything all together (zero computation :-)).

What does the rest think? I’ll make a pull request if we decide on an appropriate change.

Comments (4)