Finger Movement Classification

using a neural net

Eduard Heindl (IAP Tübingen, 1995)

Data Recording

We used a CMS^TM ultrasound based coordinate measure system with 8 markers. (^TM zebris Germany)

The makers send short ultrasonic impuls signals, which were collected by 3 micro-phones. The position (x,y,z) of the markers was then determined from the receiving time at the micros by a computer.

The repetition rate was 20 Hz.

The Angle of the Joints

After the collection of the raw data (x,y,z), we calculated the angle between the segments from the measured coordinates of the 8 markers using trigonometric equations.

Generating a Standard Pattern

To generate a standard pattern, we collected the significant part of the opening and closing of the fist.

Training of the Neuronal Net

1. Finding the two patterns x_k,x_l with the longest distance D_max in "pattern-space"

2. Initializing the Self Organizing Map (SOM) with arbitrary patterns.

3. Training of the SOM using 145 training patterns and the patterns x_k,x_l in the corner of the net.

The Resulting SOM Net

The net contains 100 patterns, which represent the most significant patterns in topological order. If it is similar to a movement of a patients with ulnar nerve palsy it is plotted in red, if there was response to each groups - grey, no response light and healthy in green color.