Here is what I am doing in research: I have a sensor that I can extract quaternion and accelerometer. My purpose is to estimate the velocity and position for short period of time(for example 20 seconds), my set my sensor frequency being 20 Hz.
The quaternion data is accurate enough since the sensor itself has built-in correction. My thought is: using quaternion info to inverse transform each accelerometer back to earth coordinate, delete gravity then integrate to get velocity and position.
But the error is huge, even I did a test I put the sensor still, the accelerometer has noise. Another is integration has drift effect. The plot of position shows the sensor is moving at long distance but the sensor actually is still.
I tried to smooth the accelerometer by averaging adjacent data, but it won't help. Anyone has any insight?
This is the core issue of using accelerometer for positioning: double integration will cumulate bias in the instrument, that may depend on the physical enviroment (e.g temperature, pressure). On the other hand, uniform noise is not an issue, since it is quickly averaged over time.
This is why there is a huge range of price in accelerometer to allow inertial navigation with some accuracy.
Otherwise Low end inertial devices like MEMS mostly intended for recognition of movement signature while gyroscopic sensors may with adequate software deliver reliable navigation in aircraft or vessels, however not the absolute position. inertial sensors may be used in hybridation with an absolute position sensor that has short period of unavailability as GPS, when driving under tunnel, or walking indoor with partial obstruction.
So I would check if characteristics of the sensor fits your application, and depending the technology, if you can calibrate it, or otherwise remediate errors, focusing on bias.
Yves Bardout Thank you Bardout. Your answer is insightful. Right now I applied high pass filter for accelerometer, and Kalman filter to velocity. The overall result is bad, but the local result is acceptable. For example, testing people is moving or not at some short time interval.