Proposed Experiments on CMU Data

Experimental Question

The primary questions wish to address in this study are:

1. Does our scale and parameter conversion method that maps between views perform appropriately on different data than upon which it was developed.

2. Can we define static body parameter measurements performed during walking that our speed indpendent.

The mechanism to explore these questions will be the CMU “Mobo” database.

Available Data and CMU Experimental design

The CMU data consists of several views of subjects on a treadmill. The data are labeled vrXX_YY for each of 3 conditions: slow walk, fast walk, and ball carrying walk. The component XX refers to view number, and YY is subject number. In particular, view 03 is from side (person walking right to left), view 05 is from about 45 degrees, and view 07 is from straight ahead.

The CMU evaluation plan as posted on their web site is the following 22 tests (“experiments”):

Exp #	Train – Slow Walk	Test – Fast Walk	Test – Ball Walk
1,2	vr03	vr03	vr03
3,4	vr03	vr05	vr05
5,6	vr03	vr07	vr07
7,8	vr05	vr03	vr03
9,10	vr05	vr05	vr05
11,12	vr05	vr07	vr07
13,14	vr07	vr03	vr03
15,16	vr07	vr05	vr05
17,18	vr07	vr07	vr07
19,20	vr03 + vr07	vr03 + vr07	vr03 + vr07
21,22	vr03 + vr07	vr05	vr05

Our initial method of performing gait recognition uses static body measurements that are sensitive to stride length. Therefore we should be able to go across views well, but poorly with different speeds. Also, our feature extraction methods will not work on frontal views. Thus any of the experiments that go between different speeds (all the odd tests) will be problematic for us.

Also, our paradigm of measuring a static body parameter feature vector does not require any specific “training” analogous to a set that helps define a model that is then applied to the gallery and the probes. When performing matching across views we do use a small number of subjects to help relate the viewing conditions. We will label them as “training” subjects..

\
To attempt to utilize the CMU data, we will first apply our algorithm to a subsets of the experiments. Those marked in bold will be first, with results by Sept 15, as they do not require new methods of feature extraction nor new methods of combining view information. They do require our development of a speed compensation method as our current procedure uses stride length as a feature and biomechanics reveals that stride length changes with speed.

Expt #	Train – Slow Walk	Test – Fast Walk	Test - Ball
1,2	vr03	vr03	vr03
3,4	vr03	vr05	vr05
5,6	vr03	vr07	vr07
7,8	vr05	vr03	vr03
9,10	vr05	vr05	vr05
11,12	vr05	vr07	vr07
13,14	vr07	vr03	vr03
15,16	vr07	vr05	vr05
17,18	vr07	vr07	vr07
19,20	vr03 + vr07	vr03 + vr07	vr03 + vr07
21,22	vr03 + vr07	vr05	vr05

We will also perform tests on the same speed but different viewing angles:

Exp #	Train – Slow Walk	Test – Slow Walk
23	vr03	vr05
24	vr03	vr07
25	vr05	vr03
26	vr05	vr07
27	vr07	vr03
28	vr07	vr05

Exp #		Train – Fast Walk	Test – Fast Walk
29		vr03	vr05
30		vr03	vr07
31		vr05	vr03
32	vr05		vr07
33	vr07		vr03
34	vr07		vr05

Exp #		Train – Ball Walk	Test –Ball Walk
35		vr03	vr05
36		vr03	vr07
37		vr05	vr03
38	vr05		vr07
39	vr07		vr03
40	vr07		vr05

MAIN PAGE