A manifold learning approach for gesture recognition from micro-Doppler radar measurements

Area of Science:

• Machine Learning
• Data Science
• Applied Mathematics

Background:

• Manifold learning is crucial for dimensionality reduction.
• Existing methods often require prior knowledge of manifold structure.
• Kernel-based approximations offer a flexible approach.

Purpose of the Study:

• To analyze the pointwise approximation error of a localized kernel method for manifold learning.
• To investigate the impact of data characteristics on approximation accuracy.
• To evaluate the method's performance and efficiency compared to deep neural networks.

Main Methods:

• Utilizing a localized kernel with least squares optimization.
• Examining error deterioration away from training data.
• Applying the method to micro-Doppler and video datasets.
• Investigating different preprocessing techniques, kernels, and manifold dimensions.

Main Results:

• Approximation error is dependent on data characteristics and proximity to training data.
• Localized kernels combined with Principal Component Analysis (PCA) components achieve near-competitive performance with deep neural networks.
• Significant improvements in training speed and memory requirements were observed.
• The method demonstrated domain-agnostic classification capabilities on video data.

Conclusions:

• The localized kernel approximation is a viable and efficient alternative for manifold learning.
• The approach offers a strong balance between performance, speed, and memory usage.
• Its domain-agnostic nature broadens its applicability across various datasets.