I'm just starting to explore topics within computer vision and curious if there are any concepts in that area that could be applied to segmenting multivariate time series with the goal of grouping individual data points similar to how a human might do the same. I know that there are a number of time series segmentation methods, but in-depth explanations of multivariate methods are more scarce and it seems like somewhat of an underdeveloped topic overall. Since segmentation is such a fundamental part of CV and is inherently multidimensional, I'm wondering if concepts there can be modified to apply to time series.
Specifically, I'd like to be able segment a time series and reformulate a prediction problem as something closer to a language processing problem. The process would look something like this:
- Segment a multivariate time series into near-homogenous segments of variable length. Some degree of preprocessing might be required but I can worry about that separately.
- Encode the properties of each segment based on summary statistics (e.g., mean, variance, derivative values, etc.) such that the segments fall into discrete buckets.
- Each bucket will represent a "word" and the goal of the model will be to predict the next word given a series of words, i.e., the next segment given a series of segments.
In a few days of reading about CV, it seems like there's a ton to learn. If there are traditional time series segmentation techniques that are more suitable, that would be of interest, but I'd still be curious about a CV approach since that approach likely better aligns with how a person might look at a graph to identify segments.