go_away

Author Topic: Feature extraction for classification of users behaviour in power mobility devic  (Read 341 times)

0 Members and 1 Guest are viewing this topic.

Offline astronautTopic starter

  • Beginner
  • *
  • Posts: 2
  • Helpful? 0
 I would like to extract some features to classify some task while my robot is performing some runs in indoor environment. I would like to know what is the state of the art and which features are relevant and how they are extracted. Feature should describe the performance of the user while driving the robot in indoor environment.Any code example regarding this?

I start with some features like linear velocity, time to accomplish the task, travelled distance., etc. . Based on this features I like to perform classification of the users run. Which methods are used for feature extraction and if possible any code example. Any help?

Thanks

Offline jwatte

  • Supreme Robot
  • *****
  • Posts: 1,272
  • Helpful? 79
To do classification, you first have to define what the axis is you want to classify.
It's impossible to answer your question before you know what it is you want to measure.
It sounds like you want to give a rating to a user -- from "A" for best to "F" for worst, for example.
But you need to know what it means to be "good" or "bad" to do this, and what's "good" and what's "bad" is very task dependent.

Offline astronautTopic starter

  • Beginner
  • *
  • Posts: 2
  • Helpful? 0
Im using power wheelchair and would like to classify the user behavior while performing some task (like 180 degree turn in some corridor) in indoor environment. I start with SVM (support Vector Machine) and used features like average linear velocity, standard deviation of orientation, idle time, traveled distance and so son. But there is no science and some fundamentals why exactly these features I used.

I get the movement data from the sensor package deployed on the wheelchair. Im classify certain action like turning 180 degree, although the user is performing that action (turning 180 degree) and Im giving him a mark (like 1 he failed, 2 he passed but his performance was bad because for example was not driving smooth, he was close to hit some obstacle etc,, 3 he was good but not perfect and 4 it was a perfect run). Im not really detecting that action because the user must perform it.

So I need some state of the art methods for feature extraction for classification of user behaviour of mobility device in indoor environment. I know that can use PCA to see the correlation of the features and to reduce the data set. Also a parametric method with some fixed or variable window size can be a good way so see which features ar relevant. But some other methods ?

To be clar Im not looking for some spatial features in the map or any localisation issues. Im trying to classify the runs of the robotics wheelchair users while they are performing some tasks in indoor environment. The tasks are part of some Assessment test by the Occupational therapist.

So I extract parameters (features) like velocity, time, distance... etc but there is no fundamentals and no validation why exactly these parameters should be used. It was only my intuition why I chose them. Furthermore, the SVM (Support Vector Machine) classify didnt performed well, the results of the classification were not good enough.

So I set like two goals. One is to improve the accuracy and the performance of the SVM (using PCA or other method) . And the second one, is to validate the method and see what is the state of the art in this problematic, which parameters (features) are more suitable and described the best the performance of the users runs. It does not have to be a wheelchair, it could be any mobile platform moving in indoor environment (at the moment trying to solve the problem indoor and later on the goal is to move outdoor).

Is it clear now? Basicly I need some example and code if posible
« Last Edit: December 05, 2013, 07:20:26 PM by astronaut »

Offline jwatte

  • Supreme Robot
  • *****
  • Posts: 1,272
  • Helpful? 79
What you need to do is to develop a model for what a "good 180 degree turn" is, and what a "bad 180 degree turn" is.
Then you can score sets of data against that model, for some measure of how good/bad that data is.
To put it another way: Before you know what you're looking for, you can't possibly find it.


 


Get Your Ad Here