How do I make a LSTM neural network for classifying motion-captured data?

Hi,

so first of all, I am pretty new to LSTM and Matlab, so not the best conditions for my request :D.

However, what I am currently trying to do is the following:

I have a sequence of motion-captured 3d-data which I want to analyse with a neural network.

I want to check if the movement that has been recorded is fullfilling some criteria. If the criteria is fullfilled, then I want to get a one as an output, else a zero.

My approach was (as I have a sequence of 3d-vectors that build up my 3d-skeleton) to give those 3d-vectors as input into a lstm-cell.

So each timestep the lstm-cell would get the current positions of each joint of the skeleton as a set of 3d-vectors.

But I am wondering what shape my output would look like, as in the end I would need a single value (0 or 1) as output.

If I understood the concept correct, then I will need a satic length of the sequences that I am trying to input into the lstm-cell. Is that correct, or is it possible to hae sequences of different lengths?

More detailed:

I have a skeleton represented by .

So as Input I thought about to make a Vector (with 81 dimensions) = X for each timestep.

(That's simply because I don't know if it is possible to use a e.g. 27 * 3 matrix as input)

So my approach was to input X for timestep T into the lstm-cell.

And to get a usable output at the end of the sequence.

Was my approach correct so far, or did I have a fundamental misunderstanding in how lstm's (or neural networks) work? Would be glad for ideas, corrections or tips, anything really that could help me.

(Also I have read some papers that classified motion-captured data by using lstm's as well. But those papers didn't really go into how the input was structurized and how the output of lstm-cells looked like.)

Update:

So i found the example of the japanese vowels in the matlab api and projected that example onto my case.

So instead of having 12 features I have 81. In the Fully-Connected-Layer instead of 9 labels / classes I only have 2.

Even if this will work I am curious on how the ouptut of the lstm directly would look like.