This paper explores the impact of different classification and feature engineering techniques in radar-based posture detection systems. The research focuses on an experimental setup involving a radar sensor used to classify human postures and the absence of human activity. The methodology includes the detection of the absence of human activity in the first stage, followed by feature engineering extraction through Variational Auto-Encoders (VAEs), and subsequent data reduction via either t-distributed Stochastic Neighbor Embedding (t-SNE) or Principal Component Analysis (PCA) if human activity is detected and posture classification is required. Five solutions differing in terms of methodology and features' dimensionality reduction strategies are presented and compared in terms of classification accuracy and computation time. The main findings reveal that all five solutions achieve classification accuracy exceeding 90%, with variations observed in computation times, showing the effectiveness of adopting VAEs to derive classification features from raw data. The research highlights a trade-off between data preservation and classification accuracy and/or computation time, demonstrating that feature extraction and data reduction techniques are a valuable enhancement to posture detection and classification.
Feature Extraction using Variational Auto-encoder for Radar-based Posture Detection Systems
2024-06-24
616762 byte
Conference paper
Electronic Resource
English
Feature-aided pose estimation approach based on variational auto-encoder structure for spacecrafts
Elsevier | 2024
|