well, the problem is that we don't know the true linear body accelerations... So one assumption is that the current attitude estimate is good enough to rotate the acceleration measured by the IMU in body frame to the LTP (LocalTangentPlane) frame. Assuming this is correct, the acceleration due to gravity is only in the z-component and b2_hff_[xy]dd_meas only contains the linear trajectory accelerations.
As Gautier already said, this was a quick'n dirty solution to some specific problems (gps lag and low accel noise) some years ago.
If you have a more powerful microcontroller (like the STM32F4 with FPU) on your autopilot, running the ahrs_mlkf (multiplicative linearized Kalman Filter in quaternion formulation) will give you better attitude estimates and hence should also improve the results of hff. But this combination hasn't really been extensively tested so far.
Keep in mind that even then hff is still not a very nice solution and if you have enough computing power a "real" INS (which estimates position, velocity, attitude and biases in one filter) will definitely give you better results. Mainly due to hardware limitations this was not implemented in Paparazzi so far.