I am trying to understand the process of sensor fusion and along with it Kalman filtering too.
My goal is to detect Fall of a device using Accelerometer and Gyroscope.
In most of the papers such as this one, It mentions how to overcome drift due to Gyroscope and noise due to Accelerometer. Eventually the sensor fusion provides us with better measurements of Roll, Pitch and Yaw and not better acceleration.
Is it possible to get better 'acceleration results' by sensor fusion and in turn use that for 'Fall detection' ? As only better Roll, Yaw and Pitch are not enough to detect a Fall.
However this source recommends to smoothen Accelerometer (Ax,Ay,Az) and Gyroscope (Gx,Gy,Gz) using Kalman filter individually and using some classification algorithm such as k-NN Algorithm or clustering to detect Fall using supervised learning.
Classification part is not my problem, it is if I should fuse the sensors(3D accelerometer and 3D gyroscope) or smoothen the sensors separately, with my goal of detecting a fall.
Several clarifications
Kalman Filter is typically to perform sensor fusion for position and orientation estimation, usually to combine IMU (accel and gyro) with some no-drifting absolute measurements (computer vision, GPS)
Complimentary filter, which is typically used to have good orientation estimation by combining accel(noisy but non-drifting) and gyro(accurate but drifting) . Using accel and combine with gyro, one can have fairly good orientation estimation. The orientation estimation you can see as primary using the gyro, but corrected using accel.
For the application of Fall detection using IMU, I believe that acceleration is very important. There is no known way to "correct" the acceleration reading, and thinking of this way is likely to be the wrong approach. My suggestion is to use accelerations as one of your inputs to the system, collect a bunch of data simulating the fall situation, you might be surprised that there are a lot of viable signals there.