I have collected 3-axis accelerometer sensor data to edge impulse project for gesture recognition. In this process, I have come across following problems:
What appropriate sample frequency should I choose for gesture recognition?
For every gesture what approriate sample length should I choose? Default: 5s
In time series data, what window size and window increase value should I choose? Default: Window size: 2000ms, Window increase: 80ms
In spectral analysis block, what appropriate values should I use for the Filter and Spectral power parameters?
For each gesture(time series data), how long shoud I collect 3-axis sensor data for training data?
I want to know is there any references or guideline to help solve these problems?
It all depends on the application. Scan the internet and search for some papers about gesture en activity recognition
For example, window size: Window Size Impact in Human Activity Recognition. This paper gives some starting points. But finally, you need to perform some experiments and check if the window you selected is the best for your application.
The best approach: start experimenting and play around with the parameters: window size, filter, … keep track of these parameters and check how this affects the accuracy of the model.
For example, build a model, keep the model fixed, Iterate through your data (features) and check the outcome. Try to understand your data and how it is related to the gesture.