Gait based gender recognition

Rohinish Gupta(rohinish[at]iitk[dot]ac[dot]in)       Roll no. 10607
Mentor:Dr. Amitabha Mukherjee(amit[at]cse[dot]iitk[dot]ac[dot]in)

Motivation

There are a number of people suffering from leg amputations across the world. The only way they can lead a normal life is through prosthetic limbs. There is a lot work which can be done in this area as the quality and functionality of knee joints available is not too much. In an effort to study the gait cycle (The combinations of knee and ankle movements which lead to walking) and the way it can be utilised to make personalised prosthetics for various people. One main classification in the style of walking is between men and women. They have a very different style of movement. Thus this project will be about detecting whether the person is a male or female by using their gait data.

Methodology

First of all a data set will be generated by taking data of different people walking on a treadmill. The data will be collected using Microsoft SDK. This data will be normalized as the experimenters will be of different build and heights. Then by using various attributes of walking like angle between the two legs and loci of movement of knee, clusters will be made for each relevant data. This clustering will be done using K means.

Then whenever new data comes, it will be tested using the above mentioned procedure and clusters will be made. By taking the mode of the results, i.e. the result which will come from every type of information, the final prediction will be generated.

References

  1. [Guillaume Garreau, Charalambos M. Andreou, Andreas G. Andreou, Julius Georgiou Salvador Dura-Bernal, Thomas Wennekers and Sue Denham] "Gait-Based Person And Gender Recognition Using Micro-Doppler Signatures".
  2. [W. Zijlstra, T. Prokop, W. Berger] "Adaptability of Leg movements during normal treadmill walking and split bent walking in children".
  3. [Presentation by Dimitris Kastaniotis, University of Patras] "Gait Recognition: Monnocular, RGB-D, Appearance and Model basedmethods".