Help me finalize my Machine learning project

I have some .json files that express the movement of a human in 3d space (each json have the 3d coordinates of every human body joint for every frame). Currently, I have 3 types of exercises and I want to classify a new movement to one of those 3 categories. The problem is that every movement can vary in size and duration and on starting and finishing point as well. My approach was to calculate features and segment every move to equal chunks to calculate the mean of each feature. So my .csv file looks like (segment1feature1,s1f2,s1,f3,..s2f1,s2f2,..,snf1,..snfn). So I had to find a way to align the moves for this approach to work. I found out about Dynamic time Wrapping algorithm and I started trying to trimming the exercises by comparing the mean of all the angles of the human body from every exercise type A with a prototype type A and every type B with a type B prototype and so on. The prototypes that I choose were short in length and clean exercises thinking that it will trim them down if I just keep only the frames that the exercise matches with the frame of the prototype exercise (by just keeping the diagonal in DTW best path line).