2
$\begingroup$

I have X, Y and Z co-ordinate of the movement patterns of a person for 30 days over some known physical layout. This is unevenly spaced time-series data with maximum frequency of 2Hz while in motion. It is known that a person can depicts one of the four patterns of locomotion i.e. lapping, pacing, random and direct. It has been defined as:

Lapping: Locomotion that has a circular path (closed loop).

Pacing: Back and forth locomotion between two end points.

Random: Locomotion along a haphazard path from source to destination.

Direct: Locomotion along a somewhat straight path from source to destination.

(Note: Source and destination are start and end of an episode. Episode is a smaller simpler navigation path which can have one of the pattern )

My aim is to find the pattern in an episode. What features I should extract from these dataset and which ML classification algorithm will be best suitable to deal with these types of problem.

I have split the movement into smaller simpler episodes where each episodes can be any of the four patters. Right now, I am using some heuristics to identify these patterns but I feel that ML can be very good to predict these patterns.

Thanks in advance :)

$\endgroup$
0
$\begingroup$

Try features in the frequency domain (fft related features), but also general statistics (mean, std, max, min,...).

An algorithm that works generally well when you don't really know which features are most significant and you have quite a few of those, is Random Forests. They're easy to train and available in most ML libraries.

This approach is commonly used for activity recognition based on IMU sensor data (accelerometer, gyroscope,...).

$\endgroup$
  • $\begingroup$ Thanks. Since I have only x,y and z co-ordinate value with time-stamp, what fft features I can calculate? $\endgroup$ – AshNTU Jul 10 '15 at 1:58
  • $\begingroup$ The first few fft components for example, calculated on a certain overlapping & sliding window (e.g. 20 coordinates with an overlap of 10, or even way more like a window of 100) $\endgroup$ – Joren Van Severen Jul 11 '15 at 8:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.