Robust Biomechanical Classification of Humans, Gorillas, and Robots under Urban Multipath Fading using Micro-Doppler Radar

Abstract

This paper presents a physically grounded framework for classifying humans, gorillas, and robots using a 35 GHz Ka-band  continuous-wave radar under realistic urban multipath conditions. Urban deployments face severe signal distortion due to multipath fading; Micro-Doppler radar is a pivotal tool for non-contact biomechanical classification. A total of twenty features are obtained from both biomechanics and spectroscopy and selected using Neighbourhood Component Analysis (NCA). Stacking ensemble learning that employs Random Forest, Support Vector Machine, K-Nearest Neighbours, and MLP, as well as linear support vector machine for meta-classifier is evaluated in five-fold stratified cross-validation using 400 original samples, augmented to 1200 using MATLAB (R2023b) . The system achieves a state-of-the-art accuracy of 90.92% ± 2.03% (95% CI: [89.14%, 92.69%]) with F1-score of 90.88% ± 2.08% under harsh urban conditions (SNR = 15 dB, K-factor=7 dB). The error analysis reveals that the use of multipath-aware augmentation and supervised feature selection leads to a significant reduction of the spectral smearing phenomenon, resulting in reduced errors for classifying human from robot at an impressive 68%. The gorilla is the class that shows the highest inter-class separability, with 95.0% accuracy. The robustness analysis demonstrates consistent behaviour across various signal-to-noise ratio and K-factor combinations due to the stacking ensemble technique that eliminates any distribution shift within the feature space. Ka- and radar detection enhances the frequency resolution of the Micro-Doppler features, improving the inter-class separability while accounting for high-order scattering phenomena. This work provides a comprehensive baseline for Micro-Doppler human-robot classification in an urban environment and it is noted that this study relies on simulation-based validation; thus, real-world Ka-band radar data is required for final empirical verification.