I'm working on an activity recognition project using smartphone sensors data and I want to implement feature extraction phase.
Assume the frequency of data collection is 20Hz and the sliding window is of size 2 sec. It means we have 40 data samples in each window. So, when we do feature extraction, we will have just one feature extracted data point for each sliding window. For example, if we use "Mean" as a feature, we should calculate the mean of all 40 data samples for each window. In fact, we have a kind of data reduction process here, from 40 raw data samples to 1 feature extracted data.
I would really appreciate it if anyone could help me which parts of my understanding are right and which parts are wrong.
Hi Meysam,
The process you described is correct. Since activities are continuous, a single signal does not contain enough information. Therefore, we divide the time into windows. 2 second is commonly used window size, but based on the target activity, you can change it.
In addition to "mean" , "max", "min", and "variance" are commonly extracted features. If you use data collected from an accelerometer or gyroscope you can generate these features also for each axis.
Good luck!
Additionally, you could use the first (and higher order) derivates to determine mean, max, min, variance of increase/decrease and for acceleration.
If you are really smart you can extract additional features from the frequency domain by FFT (Fast Fourier Transformation).
Hope this helps.
Hi Meysam,
The process you mentioned is absolutely correct. Features are extracted to represent the acquired data efficiently so as to remove the redundancy in data and obtain the equivalent information in a much more compressed form. The continuously acquired signals of sensors need to be represented by appropriate features depending upon the type of information you are seeking to enable the activity recognition process.
Hope this helps!
It is Okay, for a bit detail you may look at https://towardsdatascience.com/feature-extraction-techniques-d619b56e31be too.
A 2 sec window seems to be quite large, usually it is around 20 millisec, the logic behind this being that anything occurring in less time is not perceptible to the human auditory system and seems stationary to our ears.
Thank you SO much.
I really appreciate every single of your replies to my question.
Now I know that I'm on the right track and can further my research with more confidence. :)
Yes, your understanding of feature extraction process is absolutely right. I appreciate your thinking and dedications towards your work.
- Badges
- Science topic