Deep Learning-Based Human Activity Classification with OTFS Radar and Attention Enhanced LSTM Networks

Radar-based human activity recognition (HAR) is emerging as a resilient alternative to camera systems in industrial environments, where occlusions, reflective surfaces, and privacy concerns limit vision-based methods. This paper presents a deep learning framework that integrates convolutional feature extraction, denoising, multi-head attention, and bidirectional LSTM layers to classify activities from OTFS radar delay–Doppler signatures. Experiments in cluttered industrial environments demonstrate that denoising improves robustness against clutter, while attention stabilizes performance across longer temporal sequences. With a sequence length of 50 frames and 32 attention heads, the proposed model achieves 95.2\% accuracy on unseen test data. Visualization using t-SNE confirms clear activity separation, with minor overlap between walking and multi-person walking due to shared Doppler patterns. These results highlight the effectiveness of combining OTFS radar with attention-based temporal modeling for reliable and efficient HAR in real-world industrial monitoring.