RF-Aware CSI Sensing in Deep Neural Compression Models for Activity and Location Recognition

Abstract

Channel State Information (CSI) based RF sensing enables device free human activity recognition (HAR) and indoor location recognition, making it attractive for smart RFID-enabled and RF-assisted sensing environments. However, CSI-based sensing relies on fine-grained amplitude and phase variations that are sensitive to numerical precision, quantization noise, and edge-device constraints. Existing deep learning models for CSI processing are computationally intensive and do not explicitly account for CSI-specific signal characteristics, limiting their deployment on resource-constrained platforms. This paper presents an RF-aware and deployment-oriented deep learning framework for CSI based joint HAR and location recognition. We propose DB-XNet, a lightweight dual-branch convolutional architecture that separately processes CSI amplitude and phase, and adaptively fuses them through cross-channel attention, thereby preserving complementary RF information under resource constraints. On the ARIL benchmark, DB-XNet achieves state-of-the-art performance with 95.40% HAR accuracy and 99.30% location accuracy, while maintaining a compact footprint of only 0.54M parameters and a model size of 2.08 MB. We further perform a systematic compression aware study using multiple structured pruning methods and custom bit width quantization (2-8 bits), and validate the findings on the JUST RF CSI dataset. The results show that HAR is more sensitive to compression than location recognition, highlighting the stronger dependence of HAR on fine-grained CSI dynamics. After applying combined pruning and quantization, DB-XNet achieves up to a $5\times $ inference speedup on a Raspberry Pi 4, demonstrating its feasibility for real-time low-power RF sensing applications.