Low Complexity Radar Gesture Recognition Using Synthetic Training Data

Area of Science:

• Robotics and Automation
• Signal Processing
• Machine Learning

Background:

• Radar technology, specifically frequency-modulated continuous-wave (FMCW) radar, enables hand gesture recognition.
• Current heatmap-based methods and machine learning approaches require large datasets and complex computations.
• Collecting radar data for hand gestures is time- and energy-intensive.

Purpose of the Study:

• To propose a low computational complexity algorithm for hand gesture recognition using FMCW radar.
• To develop a synthetic hand gesture feature generator to overcome data scarcity.
• To evaluate the effectiveness of synthetic data in training radar-based gesture recognition models.

Main Methods:

• Implemented a 2D Fast Fourier Transform on radar raw data to create a range-Doppler matrix.
• Applied background modeling to isolate dynamic objects from static environments.
• Utilized Fourier beam steering for target angle calculation and Blender software for synthetic data generation.

Main Results:

• The proposed algorithm effectively separates dynamic targets from background clutter.
• Synthetic data generation using Blender provided target range, velocity, and angle information.
• An average recognition accuracy of 89.13% was achieved using synthetic data for training and real data for testing.

Conclusions:

• The developed low-complexity algorithm is efficient for radar-based hand gesture recognition.
• Synthetic data generation is a viable and effective strategy for pre-training gesture recognition models.
• This approach significantly reduces the need for extensive real-world data collection.