Does anyone have experience with recognizing single/occasional human jumps using NN and accelerometer? What would be the most suitable NN structure for that? Maybe some example from Edge Impulse close to that already exists?
I was thinking of a delay line of N accel. samples, each passed to feed-through network (thus, N inputs x # of accelerometer axes used), with 1 output - probability of jump detection.
But a single-input RNN (with internal delay elements) might also been utilized for that.
Thanks in advance for suggestions!
This should definitely be possible with a pretty standard model. We’ve done examples around workout activity detection that would be a good place to start in terms of model and DSP design:
The main difference between exercise detection and a detecting a jump is that the data taken for that exercise example was someone continuously performing that exercise. In comparison the jump will be a discrete event. Make use of the ‘split’ feature during data acquisition to separate the discrete event of the jump from the idle and other movements and avoid false classifications, as described here:
Additionally try tuning your window sizes to line up with the maximum duration you expect to see from a jump
Thanks for the fast response, David. Will give a try…