Synthetic Radar Dataset Generator for Macro-Gesture Recognition

Recent developments in mmWave technology allow the detection and classification of dynamic arm gestures. However, achieving a high accuracy and generalization requires a lot of samples for the training of a machine learning model. Furthermore, in order to capture variability in the gesture class, th...

Full description

Saved in:
Bibliographic Details
Published inIEEE access Vol. 9; pp. 76576 - 76584
Main Authors Ninos, Alexandros, Hasch, Jurgen, Alvarez, Mario Emilio Pizano, Zwick, Thomas
Format Journal Article
LanguageEnglish
Published Piscataway IEEE 2021
The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
Subjects
Animation
data creation system
Datasets
Feature extraction
Field of view
Gesture recognition
Gesture sensing
Machine learning
Millimeter waves
mm-wave technology
Model testing
Parameter modification
Radar
Radar antennas
Radar cross-sections
Solid modeling
Synthetic data
synthetic data set
Three-dimensional displays
Training
Online AccessGet full text

Cover

More Information
Summary:Recent developments in mmWave technology allow the detection and classification of dynamic arm gestures. However, achieving a high accuracy and generalization requires a lot of samples for the training of a machine learning model. Furthermore, in order to capture variability in the gesture class, the participation of many subjects and the conduct of many gestures with different arm speed are required. In case of macro-gestures, the position of the subject must also vary inside the field of view of the device. This would require a significant amount of time and effort, which needs to be repeated in case that the sensor hardware or the modulation parameters are modified. In order to reduce the required manual effort, here we developed a synthetic data generator that is capable of simulating seven arm gestures by utilizing Blender, an open-source 3D creation suite. We used it to generate 600 artificial samples with varying speed of execution and relative position of the simulated subject, and used them to train a machine learning model. We tested the model using a real dataset recorded from ten subjects, using an experimental sensor. The test set yielded 84.2% accuracy, indicating that synthetic data generation can significantly contribute in the pre-training of a model.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:2169-3536
2169-3536
DOI:10.1109/ACCESS.2021.3082843