Question summary: How can you perform a classification task on individuals through their movement/position data? I do not have a data set (in mind), but I will pretend I do in order to make my question a bit more specific:
Data example: Assume I have position data of individuals performing a certain task, e.g. doing a squat. During this activity, movement trackers were attached to 10 points of the body and their 3-dimensional data (x,y and z location) was sampled at 100 Hz. Of course, different individuals may take a different time to perform the movement. 50 individuals are professional track and field athletes, while 50 are regular people, and their labels are all known. The goal would now be to make a model which takes data of a person doing a squat, and predicts in which of the two categories they fall (binary classification).
My thoughts: There should definitely be information in the movements to improve over random guessing. However, the data is not really usable to any statistical method that I know of, such as k-means or random forests, as our (x,y,z) observations only make sense in conjunction with the rest of the movement. What we could do is compute some summarising statistics (variance in position, total squad depth etc.) for each tracker for each individual, and apply some of the mentioned methods to these. Although I think this may work to some extent, a lot of information is probably lost and the differences in the way these groups would move may be very subtle and hard to capture by such summarising statistics.
I am curious if another approach is possible, i.e. if in some way you can use the full data. For example feed it in a neural network in some way and have it sort itself out to get a sensible result, or if even with way more data that just doesn't make sense. If it doesn't, then I am wondering how you could even approach such a summarising statistic extraction (feature extraction? when using that term I think of e.g. PCA, that doesn't seem to apply), in this specific example of movement data that seems hard, the mean of the x-position makes little sense to use, for example.
I have seen two somewhat similar questions which helped a bit. One has similar data but is about classifying movements, but this person asked very specific questions and I couldn't make out how they manipulated the data right now to analyse, link. In addition, I saw a question which was very similar in the sense that you see a time series with values and then have to make a binary classification, link. The answers here purely mentioned feature extraction, which makes me think that this is the only approach that can be taken (but I am asking to make sure), but even then it was not clear to me how to apply feature extraction to movement data, as classical summarising metrics may make little sense. Is there any general approach here?
Apart from answering the question about the example data, I would appreciate any good literature/sources on similar problems, especially from a mathematician's perspective. I do not have a specific task I want to perform, but am quite curious how these problems can be approached.
