WiDFS 3.0: Lightweight SISO Sensing for ISAC

• WiDFS 3.0 is a bistatic SISO sensing framework for ISAC that employs a two-stage pipeline combining self-referencing cross-correlation and delay-domain beamforming to resolve phase distortions and Doppler ambiguity.
• It uses self-referencing cross-correlation to remove random phase offsets from CSI and delay-domain beamforming to extract unambiguous delay-Doppler-time features, enabling robust sensing with a single antenna.
• Designed for low-complexity, real-time deployment on IoT and edge devices, WiDFS 3.0 achieves high-accuracy delay estimation and efficient feature extraction, outperforming conventional multi-antenna methods in challenging environments.

Searching arXiv for WiDFS 3.0 and related WiDFS papers to ground the article in the cited literature. WiDFS 3.0 is a lightweight bistatic Single-Input Single-Output (SISO) sensing framework for integrated sensing and communication (ISAC) that enables accurate delay and Doppler estimation from distorted Channel State Information (CSI) by suppressing Doppler mirroring ambiguity, while operating with only a single antenna at both the transmitter and receiver (Wang et al., 18 Aug 2025). It is designed for low-complexity deployments in which clock asynchrony introduces random phase offsets that cannot be mitigated by conventional multi-antenna methods. Its central technical contribution is a two-stage pipeline: self-referencing cross-correlation (SRCC) for SISO random phase removal, followed by delay-domain beamforming to resolve Doppler ambiguity and produce unambiguous delay-Doppler-time features for compact neural networks (Wang et al., 18 Aug 2025).

1. Lineage, scope, and nomenclature

The name WiDFS originates in the earlier WiFi Doppler Frequency Shift framework, which addressed single-target real-time passive tracking using CSI collected from commercial-off-the-shelf WiFi devices. That earlier system assumed a transmitter with a single antenna and a receiver with three antennas, removed transceiver asynchronization by CSI cross-correlation between RX antenna pairs, estimated a Doppler frequency shift in a short-time window, separated dynamic human components from CSI self-correlation terms, and then separately calculated angle-of-arrival and human reflection distance for tracking; a prototype reported a median position error of 72.32 cm in multipath-rich environments (Wang et al., 2021). WiDFS 3.0 preserves the emphasis on physically grounded signal processing, but shifts the operating point to bistatic SISO sensing, where cross-antenna cancellation is unavailable and Doppler mirroring becomes a first-order problem (Wang et al., 18 Aug 2025).

The “3.0” suffix denotes a version within this wireless sensing line rather than a connection to Web 3.0 or to LLMs with similar versioning. In the cited literature, “Web 3.0” refers to a semantic-web and blockchain-centered internet architecture (Fan et al., 2023), while “WiNGPT-3.0” denotes a 32-billion-parameter medical reasoning model (Zhuang et al., 23 May 2025). WiDFS 3.0 instead belongs to the ISAC and CSI-based sensing literature (Wang et al., 18 Aug 2025).

2. Bistatic SISO sensing problem and design objectives

WiDFS 3.0 is motivated by the observation that real-world deployment is often limited to low-cost, single-antenna transceivers. In such a bistatic SISO setup, clock asynchrony introduces random phase offsets in CSI, and these offsets cannot be mitigated using conventional multi-antenna methods. The framework therefore targets three coupled challenges: random phase distortion caused by timing and frequency offsets, Doppler mirror ambiguity in inferred motion direction, and the need for features and models that remain robust under user, device, and environment variability while remaining deployable on modest hardware (Wang et al., 18 Aug 2025).

The framework is explicitly positioned as a practical, real-time ISAC pipeline. It aims to extract ambiguity-resolved, interpretable delay and Doppler information exclusively from SISO CSI, then package that information into compact learning-ready representations. For single-target tasks, the delay dimension can be summed or compressed into a Doppler-time map; for multi-target tasks, the full three-dimensional delay-Doppler-time tensor is retained and fed directly to lightweight neural networks such as MobileViT-XXS and MobileNetV2 (Wang et al., 18 Aug 2025). This suggests an architectural preference for preserving physically meaningful structure before statistical learning.

3. Self-referencing cross-correlation (SRCC)

SRCC is the mechanism by which WiDFS 3.0 removes random phase effects without relying on line-of-sight assumptions or multi-antenna diversity. The procedure begins with delay-domain CSI reconstruction. For each per-symbol CSI vector, an inverse Fourier transform is applied to obtain the delay-domain channel impulse response: $$h_j\left(\tau\right) = \mathcal{F}^{-1} \{ \mathit{CSI}_{i,j} \} = \sum_{i=0}^{N-1} \mathit{CSI}_{i,j} e^{j2\pi f_i \tau}.$$ A Gaussian window centered at the dominant energy bin $\tau^{\text{peak}_j}$ is then used to enhance the dominant path and suppress weaker multipath components: $${h}^{'}_j(\tau) = \mathcal{G}(\tau - \tau^{\text{peak}_j}) \cdot h_j(\tau), \quad\mathcal{G}(\tau) = \exp\left(-\left(\frac{\tau}{2\sigma}\right)^2\right).$$ The result is transformed back to the frequency domain,

$$\mathcal{CSI}_{i,j} = \mathcal{F} \{ {h}^{'}_j(\tau) \} = \sum_{\tau=0}^{N-1} {h}^{'}_j(\tau) e^{-j2\pi f_i \tau},$$

and the original CSI is cross-correlated with this reconstructed reference: $$\Delta \mathit{CSI}_{i,j} = \mathit{CSI}_{i,j} \cdot \overline{\mathcal{CSI}_{i,j}}.$$ According to the paper, this removes the common phase offsets due to TO/CFO and hardware while preserving the relative phase structure needed for delay and Doppler estimation (Wang et al., 18 Aug 2025).

The Gaussian window width $\sigma$ governs a stated trade-off. Wider windows preserve more energy and reduce phase noise, but may admit more multipath; narrower windows better suppress multipath but can reduce SNR. The paper further gives a CRLB-style bound on phase estimation variance: $$\mathrm{Var}(\phi_{i,j}) \geq \frac{\eta^2}{\left\| \mathcal{G}\left(\tau - \tau_j^{\text{peak}}\right) \cdot h_j\left(\tau\right) \right\|^2}.$$ This formulation situates SRCC as a low-cost compensation method based only on IFFT, FFT, and elementwise products, with no need for additional antennas or any prior CSI calibration (Wang et al., 18 Aug 2025).

4. Delay-domain beamforming and ambiguity-resolved features

After phase compensation, WiDFS 3.0 resolves Doppler ambiguity in the delay domain rather than the spatial domain. The compensated CSI is averaged across time to estimate a static background,

$$\mathcal{U}_i = \frac{1}{M} \sum_{j=1}^{M} \Delta \mathit{CSI}_{i,j},$$

and the dynamic component is obtained by subtraction,

$$\mathcal{V}_{i,j} = \Delta \mathit{CSI}_{i,j} - \mathcal{U}_i,$$

followed by normalization,

$$\mathcal{W}_{i,j} = \frac{\mathcal{V}_{i,j}}{\mathcal{U}_i}.$$

An observation matrix is then built by concatenating the normalized signal and its conjugate: $\tau^{\text{peak}_j}$0 This construction is used to exploit conjugate symmetry in the signal model (Wang et al., 18 Aug 2025).

For each candidate relative delay, WiDFS 3.0 defines a delay-domain steering vector

$\tau^{\text{peak}_j}$1

and computes MVDR beamforming weights using forward-backward smoothed covariance: $\tau^{\text{peak}_j}$2

$\tau^{\text{peak}_j}$3

For each delay bin, a temporal Fourier transform yields the Doppler spectrum,

$\tau^{\text{peak}_j}$4

Stacking over delays gives a two-dimensional delay-Doppler spectrum; stacking across multiple coherent processing intervals gives a three-dimensional delay-Doppler-time feature (Wang et al., 18 Aug 2025).

The paper attributes two main benefits to this stage. First, delay-domain beamforming suppresses Doppler mirror artifacts even in SISO. Second, it suppresses by-product noise terms by exploiting conjugate symmetry. A common misconception in this area is that bistatic SISO CSI is intrinsically limited to mirrored Doppler features; WiDFS 3.0 argues instead that ambiguity can be resolved through signal-model-aware processing in the delay domain (Wang et al., 18 Aug 2025).

5. Experimental setting and empirical performance

The reported experiments use Intel 5300 WiFi NICs at 5 GHz and USRP for LTE at 3.1 GHz, both in SISO operation, and include indoor line-of-sight and non-line-of-sight environments with both single-target and multi-target scenarios. The datasets include Widar 3.0 for HCI gestures and digits, with approximately 200k samples, WiMANS for multi-user activity, WiDFS, and LTE datasets. Runtime platforms include Raspberry Pi 4B (8GB) and MacBook Pro. Typical parameter settings are a CPI of 128 OFDM symbols, Gaussian window $\tau^{\text{peak}_j}$5 with IFFT/FFT bin size 128, 32 delay bins for WiFi at 20 MHz and 1 m resolution, and a Doppler FFT with 128 bins over $\tau^{\text{peak}_j}$6 Hz (Wang et al., 18 Aug 2025).

Under this setup, SRCC is reported to achieve the lowest median range estimation error, at 2.02–2.16 meters at the 50% percentile and 3.22–3.42 m at the 70% percentile, even outperforming multi-antenna DCACC and other SISO methods such as CASR and CFCC. The paper also states that Doppler ambiguity is significant for CACC and regular SISO approaches, whereas SRCC with delay-domain beamforming yields clear, unambiguous Doppler profiles and micro-Doppler spectra that are visually and quantitatively comparable to multi-antenna DCACC (Wang et al., 18 Aug 2025).

For downstream recognition, the extracted features are evaluated with compact neural networks. On Widar 3.0, MobileViT-XXS with SRCC features attains an F1 of 0.928–0.938, compared with BVP’s 0.849–0.859 with multi-receiver features. On WiMANS, MobileViT-XXS plus SRCC achieves F1 of 0.659 for activity recognition and 0.629 for people counting. The paper further reports that, when deployed on the embedded-friendly MobileViT-XXS with only 1.3M parameters, WiDFS 3.0 consistently outperforms conventional features such as CSI amplitude, mirrored Doppler, and multi-receiver aggregated Doppler. Mean feature extraction latency is 8.5 ms on Raspberry Pi 4B without code optimization, which the authors describe as enabling true real-time operation (Wang et al., 18 Aug 2025).

6. Significance, misconceptions, and future directions

WiDFS 3.0 is presented as the first practical SISO bistatic solution with no need for multi-antenna, line-of-sight, or high-computation methods, and as the first demonstration that SISO bistatic ISAC can offer ambiguity-resolved, high-accuracy sensing features rivaling or besting more complex hardware approaches (Wang et al., 18 Aug 2025). Within the trajectory from the original WiDFS system to WiDFS 3.0, the conceptual shift is from multi-antenna passive tracking toward single-antenna ambiguity resolution, while retaining an emphasis on explicit signal models rather than purely end-to-end feature learning (Wang et al., 2021).

The paper’s claims are strongest in two areas: delay estimation, where SISO performance is reported as comparable to or even surpassing that of prior multi-antenna methods, especially in delay estimation, and embedded deployment, where the combination of compact signal processing and small models targets low-cost IoT and edge platforms (Wang et al., 18 Aug 2025). A plausible implication is that WiDFS 3.0 narrows the usual trade-off between interpretability and deployability: its feature tensors are physically grounded, yet they are also compact enough for lightweight neural inference.

The framework is also described as scalable and extensible. When multi-antenna hardware becomes available, the same line of work suggests extension to SIMO or MIMO for AoA and AoD estimation (Wang et al., 18 Aug 2025). More broadly, WiDFS 3.0 contributes to the view that the limitations of SISO sensing are not exhausted by raw hardware constraints; they also depend on whether phase compensation and ambiguity suppression are formulated in a way that exploits the latent structure of CSI.