0

I am working on a project that involves the comparison of motion data collected from accelerometer and gyroscope sensors. Let's say that i have data for a specific type of motion (e.g throwing ball) and i have manually created segments as you can see in the image bellow, where A B and C indicate three segments.

Accelerometer data with segments: accelerometer data with segments

Goal: Given X a new time series of motion data that i recorded from the same sensor how can i tell how similar the motion was for each segment with motion data Y from my collection. Note that for the new data i don't have the information about segments.

Ideas: I know how to compare two time series with cross correlation or DTW but in my case there is a problem with the segments. Is it a good idea to apply DTW with 1-KNN to find the 'nearest' time series (of X) from my collection and then compare each segment of Y_nearest with X?How can i do the comparison if i don't have the information about segments for X motion data? Is there a better approach using machine learning algorithms?

Radiodef
  • 37,180
  • 14
  • 90
  • 125
kostilatos90
  • 11
  • 4

1 Answers1

0

The data here is just on the broader of not being amiable to DTW or other shape measures, and requiring a feature based method.

DTW can align misaligned data, but it cannot map say 5 peaks to 4 peaks. For such data, people resort to creating a feature vector of {min,max,STD, zero-crossings, etc}

However, there is an novel unpublished distance measure that might work. If you write to me I will send you the code.

Eamonn Keogh

eamonn
  • 91
  • 1