Modelling of Hand Gestures and Motions

Abstract

Measuring the radar response has proven to be a valid input for neural networks
used for gesture and motion recognition. Previous work have been based on measuring
the radar response of hand gestures, but it has been shown that models
based on the analytic expressions for basic objects are useful when simulating
larger motions, such as pedestrians. This work explores if a hand model made of
basic objects can reach the same results as actual hands. To do this, this basic
objects model is implemented in two different solvers to compare the difference
in value and run-time between analytic expressions and Finite Element Method
solutions. The Finite Element solution is shown to change value based on the simulated
gesture, but the analytic model is proving to be far faster. In the end, the
latter is not capable of reproducing the response of all motions which have been
measured. Implementing object interaction into the analytic model should give
this model the details required to simulate even the smaller gestures. Therefore,
this model is still work in progress.

Popular Abstract

The initial use for radar-technology was for long-range airplane detection, but its
uses have branched off into a different purpose. This new branch is instead based
on short-range detection. One major part of this branch is the use of radar in traffic
situations. With the availability of this short-range motion detection, the usability
for radar in motion and gesture recognition has been explored. A contact-free,
motion-based system control has long been a goal, and previous steps have been
efficient at motion capture but inefficient when considering power consumption.
Based on the success of previous work done on radar-based gesture recognition,
the following work focuses on the possibility of simulating accurate gestureresponse.
Previous studies have shown that radar-based gesture recognition is
possible by using neural networks. To make these neural networks able to do
what they are capable of, training-data is needed. One method for obtaining
this training-data is to record individuals performing motions and gestures. This
would be strenuous and tedious, which is why simulated data would be of great
assistance. What if simulated data could be calculated close enough to actual
measurements? To answer this, this work will examine one method of simulating
radar response.
The simulations will be based on analytical expressions for the radar cross
section (RCS) of three basic objects. One plate, fourteen spheres, and fourteen
cylinders are used to represent the structure of the hand, so how well these objects
are simulated will first be controlled. This is done by implementing a plate, a
sphere, and a cylinder, and then first examining the frequency-response of the
formulas for the individual objects. The objects, which will be simulated as made
of perfectly electrically conductive (PEC) materials, will then be compared to
results from a Physical Optics (PO) approximation. And, lastly, the inherent
implementation of a time delay will be examined so that the objects placed at a
distance from the radar will be simulated to be at said distance. The important
reason as to why a model based on analytic expressions would be preferable to
a model based on the physical optics approximation, is that the physical optics
model would take far longer time. This will therefore be examined for both the
comparison of the simulations for the basic objects, and also in the next step when
a hand is simulated.
After the objects have been examined and made sure to have been implemented
properly, a model of a hand will then be implemented. This hand will describe the positions and directions of the objects. This will be done by first defining
the position and direction of the palm. The fingers and the thumb will then be
implemented by defining the location of the first joint of each in relation to the
position and direction of the palm. Then the first phalanx of each finger will
be implemented based on the position of the first joint and the direction of said
phalanx. In the end, this leads to that the whole hand can be defined by using
the location and direction of the palm and the angle of each of the joints as input.
The implemented hand will then be examined and compared to physical optics
approximation results from two gestures, and, lastly, four gestures will then be
compared to measured data.
With this analytical model it will be shown that the response of different gestures
can be simulated. However, the way that the objects have been implemented
means that the objects do not interact, and the simulated data is therefore not
similar enough to be used as training data.

• author: Ingemarsson, Erik
• supervisor: Lars Ohlsson; Sebastian Heunisch
• organization: Department of Electrical and Information Technology
• alternative title: Modellering av Handgester och rörelser
• course: EITM01 20191
• year: 2019
• type: H2 - Master's Degree (Two Years)
• subject: Technology and Engineering
• keywords: Radar, Gesture Recognition, Signal Processing
• report number: LU/LTH-EIT 2019-713
• language: English
• id: 8990313
• date added to LUP: 2019-07-12 10:24:19
• date last changed: 2019-07-12 10:24:19