Single-Channel Radar Indoor Localization

• The paper demonstrates that single-channel FMCW radar can achieve decimeter-scale indoor localization by leveraging passive reflector measurements and range-Doppler processing.
• Methodologies combine fingerprinting, channel-charting, and distributed SISO radar configurations to extract range, Doppler, and multipath signatures under challenging NLoS conditions.
• Empirical results show RMSE values as low as 13–20 cm, highlighting the effectiveness of hardware minimalism and advanced multipath exploitation in indoor positioning.

Searching arXiv for recent and foundational papers on single-channel or single-radar indoor localization, channel-charting, passive reflectors, and radar scene understanding. Single-channel radar-based indoor localization denotes a class of indoor positioning systems that deliberately avoid full multi-channel angle-imaging hardware and instead infer position from range, Doppler, amplitude, multipath structure, or engineered environmental references. In its strictest usage, the term refers to a single-channel FMCW radar on a mobile platform observing passive reflectors at known locations, so that localization is achieved with one receive channel and low hardware complexity (Hinderer et al., 19 Sep 2025, Hinderer et al., 19 Sep 2025). In adjacent usages, closely related goals are pursued with multiple distributed SISO radars (Hamidi et al., 2020, Qin et al., 17 Apr 2026), a single installed radar node that is not literally single-channel because it relies on MIMO angle estimation (Harisha et al., 7 Jul 2025, Zhou et al., 18 Nov 2025), or single-anchor radio/UWB systems that use channel impulse response fingerprints rather than explicit geometric inversion (Li et al., 2022, Stahlke et al., 2022). The topic is therefore unified less by one hardware definition than by one systems objective: indoor localization with minimal infrastructure and without dependence on dense optical sensing, especially under non-line-of-sight (NLoS) conditions.

1. Scope, terminology, and architectural variants

The term is technically heterogeneous. Some works use “single-channel” literally, meaning a single receive chain or a single-channel FMCW radar (Hinderer et al., 19 Sep 2025, Hinderer et al., 19 Sep 2025). Others use “single” to mean a single radar installation / single node / single anchor, even when multi-antenna MIMO processing is essential, as in the mmWave backscatter system "DragonFly: Single mmWave Radar 3D Localization of Highly Dynamic Tags in GPS-Denied Environments" (Harisha et al., 7 Jul 2025). A third group concerns single RF chain per side rather than monostatic radar, as in "An Indoor Environment Sensing and Localization System via mmWave Phased Array" (Sun et al., 2022). A fourth group is only partially transferable: multi-anchor synchronized SISO channel-charting systems (Stahlke et al., 2022) and single-anchor UWB CIR-fingerprinting systems (Li et al., 2022).

The main architectural families can be organized as follows.

Family	Core sensing setup	Relation to single-channel radar
Passive-reflector FMCW IPS	Single-channel FMCW radar plus passive reflectors	Strict match
Distributed SISO radar	Multiple monostatic SISO nodes	Single-channel per node
Single-node MIMO radar	One radar installation with multiple Tx/Rx	Single node, not strict single-channel
Single-anchor CIR localization	One anchor, multipath fingerprinting	Radar-analogous, not radar proper

The strict single-channel formulation is the most hardware-minimal. A mobile robot carries one upward-looking FMCW radar, while the environment contains passive local reference points mounted at known positions above the robot (Hinderer et al., 19 Sep 2025, Hinderer et al., 19 Sep 2025). Because a single channel does not directly provide angle of arrival, localization is transferred to the environment: geometry is encoded by which reflectors are visible and at what ranges.

A common misconception is that any “single radar” system is automatically a single-channel system. This is false in the present literature. "DragonFly" uses a 2-Tx, 8-Rx time-division MIMO FMCW radar and derives 3D position from range, azimuth, and elevation; it is therefore a one-node system but not a one-channel system (Harisha et al., 7 Jul 2025). Likewise, RISE is a single static radar benchmark, but its sensing model is built on MIMO beamforming and separate AoA/AoD processing (Zhou et al., 18 Nov 2025). By contrast, the passive-reflector AMR systems explicitly target one receive channel and treat unknown reflector correspondence as the central inference problem (Hinderer et al., 19 Sep 2025, Hinderer et al., 19 Sep 2025).

2. Measurement models and physical observables

In strict single-channel FMCW radar, the fundamental observable is the real-valued baseband matrix

$\mathbf{X}\in\mathbb{R}^{M\times N},$

with elements

$$x\left(m,n\right) = \sum_{k=1 }^{K}A_k\cos\left(-2\pi\left(\frac{2Br_k}{T_{chirp}c}T_m m+\frac{2f_0 v_k}{c}\left(T_m m+T_n n\right)+\phi_k\right)\right)+\nu(m,n),$$

where $r_k$, $v_k$, and $A_k$ are the range, radial velocity, and amplitude of the $k$-th reflection, and $\nu(m,n)$ is additive white Gaussian noise (Hinderer et al., 19 Sep 2025). Range extraction uses the range FFT

$$S(o,n) = \sum_{m=0}^{M-1} x(m,n) e^{-j 2\pi \frac{om}{M}},$$

with power profile $R(o,n)$, while Doppler estimation is obtained by a second FFT over chirps to form $T(o,p)$ and the range-Doppler map $$x\left(m,n\right) = \sum_{k=1 }^{K}A_k\cos\left(-2\pi\left(\frac{2Br_k}{T_{chirp}c}T_m m+\frac{2f_0 v_k}{c}\left(T_m m+T_n n\right)+\phi_k\right)\right)+\nu(m,n),$$0 (Hinderer et al., 19 Sep 2025). In the passive-reflector systems, the localization observables are explicitly range, Doppler, and amplitude of detected local reference points (Hinderer et al., 19 Sep 2025).

Indoor radar localization is fundamentally multipath-limited, and the most rigorous theoretical treatment in this corpus models the received signal as

$$x\left(m,n\right) = \sum_{k=1 }^{K}A_k\cos\left(-2\pi\left(\frac{2Br_k}{T_{chirp}c}T_m m+\frac{2f_0 v_k}{c}\left(T_m m+T_n n\right)+\phi_k\right)\right)+\nu(m,n),$$1

where the deterministic part is a sum of resolvable specular multipath components, the diffuse multipath term is $$x\left(m,n\right) = \sum_{k=1 }^{K}A_k\cos\left(-2\pi\left(\frac{2Br_k}{T_{chirp}c}T_m m+\frac{2f_0 v_k}{c}\left(T_m m+T_n n\right)+\phi_k\right)\right)+\nu(m,n),$$2, and $$x\left(m,n\right) = \sum_{k=1 }^{K}A_k\cos\left(-2\pi\left(\frac{2Br_k}{T_{chirp}c}T_m m+\frac{2f_0 v_k}{c}\left(T_m m+T_n n\right)+\phi_k\right)\right)+\nu(m,n),$$3 is noise (Leitinger et al., 2014). In a floor-plan-aware formulation, each specular path is associated with a virtual anchor, and the delay is written as

$$x\left(m,n\right) = \sum_{k=1 }^{K}A_k\cos\left(-2\pi\left(\frac{2Br_k}{T_{chirp}c}T_m m+\frac{2f_0 v_k}{c}\left(T_m m+T_n n\right)+\phi_k\right)\right)+\nu(m,n),$$4

The corresponding equivalent Fisher information matrix formalizes which multipath components actually carry usable position information and when path overlap destroys that information (Leitinger et al., 2014). For monostatic radar, this is directly relevant: some reflections yield strong geometric sensitivity, while others can have vanishing delay gradient and contribute no localization information.

Adjacent single-anchor and channel-fingerprinting approaches replace explicit geometric inversion by learned statistics of impulse responses. "Single-anchor UWB Localization using Channel Impulse Response Distributions" models the magnitude vector $$x\left(m,n\right) = \sum_{k=1 }^{K}A_k\cos\left(-2\pi\left(\frac{2Br_k}{T_{chirp}c}T_m m+\frac{2f_0 v_k}{c}\left(T_m m+T_n n\right)+\phi_k\right)\right)+\nu(m,n),$$5 with an area-conditioned multivariate Gaussian mixture $$x\left(m,n\right) = \sum_{k=1 }^{K}A_k\cos\left(-2\pi\left(\frac{2Br_k}{T_{chirp}c}T_m m+\frac{2f_0 v_k}{c}\left(T_m m+T_n n\right)+\phi_k\right)\right)+\nu(m,n),$$6, and classifies a snapshot by

$$x\left(m,n\right) = \sum_{k=1 }^{K}A_k\cos\left(-2\pi\left(\frac{2Br_k}{T_{chirp}c}T_m m+\frac{2f_0 v_k}{c}\left(T_m m+T_n n\right)+\phi_k\right)\right)+\nu(m,n),$$7

or a short set by the similarity score

$$x\left(m,n\right) = \sum_{k=1 }^{K}A_k\cos\left(-2\pi\left(\frac{2Br_k}{T_{chirp}c}T_m m+\frac{2f_0 v_k}{c}\left(T_m m+T_n n\right)+\phi_k\right)\right)+\nu(m,n),$$8

This makes the full delay-profile shape, rather than a single range estimate, the basic localization observable (Li et al., 2022). "Indoor Localization with Robust Global Channel Charting" formalizes an analogous principle for synchronized SISO radio links by requiring

$$x\left(m,n\right) = \sum_{k=1 }^{K}A_k\cos\left(-2\pi\left(\frac{2Br_k}{T_{chirp}c}T_m m+\frac{2f_0 v_k}{c}\left(T_m m+T_n n\right)+\phi_k\right)\right)+\nu(m,n),$$9

then approximating the local distance with aligned CIR magnitudes and learning a 2D chart through geodesic distances and a Siamese network (Stahlke et al., 2022). This suggests that strict single-channel radar systems can also be framed as learning problems over structured time-delay signatures, not only as explicit multilateration problems.

3. Localization paradigms and representative system designs

The strict passive-reflector formulation is the clearest low-complexity architecture. In "Indoor Positioning Based on Active Radar Sensing and Passive Reflectors: Reflector Placement Optimization" the radar is mounted on the AMR at height $r_k$0, reflectors are at known positions $r_k$1, and the room is discretized into a planar grid (Hinderer et al., 19 Sep 2025). Global localization uses a fingerprint map

$r_k$2

built from the nearest $r_k$3 visible reflectors, their rounded ranges, and reflector types. Local refinement is evaluated through the geometry matrix

$r_k$4

and GDOP

$r_k$5

The operational workflow is fingerprinting for initialization, AMCL for recursive state estimation, then multilateration once reflector association becomes tractable (Hinderer et al., 19 Sep 2025).

"Hybrid Baseband Simulation for Single-Channel Radar-Based Indoor Localization System" adds a signal-level simulator for the same architectural concept (Hinderer et al., 19 Sep 2025). Its central contribution is hybridization: room multipath and clutter are generated by Matlab ray tracing, passive reference reflectors are modeled separately through FEKO RL-GO radar cross-section simulation, and the real radar’s measured two-way gain is inserted into the baseband synthesis. This enables end-to-end simulation of detection, false alarms, missed detections, and particle-filter localization under realistic clutter (Hinderer et al., 19 Sep 2025).

A second paradigm uses multiple distributed single-channel radars rather than one. "Single Channel MMWave FMCW Radar for 2D Target Localization" combines two low-cost 79 GHz single-channel radars through bilateration:

$r_k$6

then suppresses ghost targets using CFAR, pruning, and multi-target tracking with a constant-velocity Kalman filter and Global Nearest Neighbor association (Hamidi et al., 2020). "Multi-site Radar Systems for High-Precision Indoor Positioning and Tracking" generalizes this distributed-SISO idea and introduces velocity synthesis-assisted localization. Each radar measures range and radial velocity,

$r_k$7

with

$r_k$8

and candidate positions are retained only if they yield mutually consistent synthesized velocities across radar pairs and across short temporal intervals (Qin et al., 17 Apr 2026). This replaces MIMO angle estimation by cross-site Doppler geometry.

A third paradigm uses a single radar node with richer angular processing, which is adjacent rather than strict. "DragonFly" uses intra-chirp backscatter modulation and TDM-MIMO phase processing to recover range, azimuth, and elevation from a single installed radar, but this depends critically on multiple Rx channels, multiple Tx channels, and a custom tag (Harisha et al., 7 Jul 2025). RISE likewise uses a single static radar node but exploits a full Range–AOA–AOD cube,

$r_k$9

to recover off-diagonal multipath and infer hidden reflectors from ghost geometry (Zhou et al., 18 Nov 2025). These works demonstrate what one-node systems can do once angular diversity is reintroduced.

4. Multipath exploitation, map learning, and unsupervised geometry recovery

A defining trait of indoor radar localization is the transition from multipath suppression to multipath exploitation. The estimation-theoretic framework of "Evaluation of Position-related Information in Multipath Components for Indoor Positioning" shows that resolvable specular multipath components behave as deterministic virtual-anchor measurements, while diffuse multipath acts as colored interference (Leitinger et al., 2014). In the non-overlapping case, the delay information of path $v_k$0 scales as

$v_k$1

so waveform bandwidth, path SINR, and path resolvability jointly determine localization quality (Leitinger et al., 2014). This establishes why high bandwidth and explicit path management are indispensable in single-channel radar.

RISE takes a more constructive route. It detects a moving human $v_k$2 and first- and second-order ghost reflections $v_k$3, then inverts the ghost geometry to estimate reflector locations. For first-bounce ghosts, it uses

$v_k$4

followed by a reflector-point reconstruction in the direction of the ghost return (Zhou et al., 18 Nov 2025). The resulting sparse reflector points are clustered with GMM, fitted with RANSAC, and completed by a Sim2Real Hierarchical Diffusion model. This suggests a path from raw single-radar multipath to room maps that can later constrain localization, even when direct wall visibility is incomplete.

Channel-fingerprinting approaches provide a complementary, label-efficient route. "Indoor Localization with Robust Global Channel Charting" shows that dense unlabeled SISO measurements can be embedded into a 2D channel chart by approximating geodesic distances with a Siamese neural network and then aligning that chart to world coordinates by a simple affine transform (Stahlke et al., 2022). "Single-anchor UWB Localization using Channel Impulse Response Distributions" shows that modeling the joint distribution of multiple delay bins and aggregating evidence over short CIR sets markedly improves area classification over treating each delay bin independently (Li et al., 2022). These methods are not radar systems in the strict sense, but they are directly relevant wherever a single radar channel produces stable, location-dependent range-profile or impulse-response signatures.

5. Empirical performance across the literature

Strict single-channel reflector-based systems presently offer the clearest decimeter-scale evidence. In the passive-reflector optimization study, fingerprinting with one reflector type achieved $v_k$5 for a random 32-reflector layout and $v_k$6 for an optimized 24-reflector layout; the paper emphasizes that $v_k$7 cm is less than twice the radar range resolution $v_k$8 cm (Hinderer et al., 19 Sep 2025). The hybrid baseband simulation paper reports a prior particle-filter localization result from the earlier high-level simulator with mean RMSE around 13 cm, but does not yet provide a full end-to-end localization benchmark using the new signal-level simulator (Hinderer et al., 19 Sep 2025).

Distributed SISO systems show that multiple single-channel nodes can close much of the gap to MIMO. The multi-site VSA system reports simulation RMSE improvement from 0.12 m to 0.06 m relative to trilateration, and real single-target tracking errors of 10.5 cm, 14.0 cm, and 16.3 cm on rhombus, circle, and star trajectories, outperforming single-MIMO-radar and point-cloud-fusion baselines (Qin et al., 17 Apr 2026). By contrast, the older two-radar bilateration paper demonstrates successful 2D localization qualitatively, but does not report a rigorous numerical localization RMSE in the extracted text (Hamidi et al., 2020).

Adjacent one-node MIMO systems reach higher accuracy under stronger hardware assumptions. "DragonFly" reports 12 cm median total 3D error for dynamic backscatter tags, with median errors of 6 cm in range, 1.4° in azimuth, and 0.38° in elevation, but this depends on a 2-Tx, 8-Rx MIMO FMCW radar, custom tags, and Doppler disambiguation (Harisha et al., 7 Jul 2025). RISE reports a Chamfer Distance of 16.03 cm and F1-score of 83.63 for layout reconstruction, plus 57.78% IoU for object detection, showing that a single installed radar can reconstruct geometry well enough to support later localization tasks, provided motion-induced multipath is available (Zhou et al., 18 Nov 2025).

Single-anchor and unsupervised fingerprinting systems remain competitive in NLoS-heavy positioning. Robust global channel charting reports localization accuracies of 0.69 m for the UWB and 1.4 m for the 5G setup after affine alignment (Stahlke et al., 2022). Single-anchor UWB CIR-distribution modeling reaches 100.0% area classification in the classroom A1→A2 condition with MD-GMM-MaxSim and 95.0% in computer-room NLoS A1→A2 with MD-GMM-SVC, confirming that full multipath fingerprints can be highly discriminative even without trilateration (Li et al., 2022).

6. Assumptions, limitations, and research directions

The central limitation of strict single-channel radar is observability. A lone fixed SISO radar measures range and radial velocity but does not directly recover cross-range position. This is why passive reflectors, motion-induced diversity, floor-plan priors, or multiple distributed single-channel nodes recur throughout the literature (Hinderer et al., 19 Sep 2025, Qin et al., 17 Apr 2026, Leitinger et al., 2014). This suggests that strict single-channel systems are best understood as range/Doppler-and-map systems rather than direct range-angle imagers.

A second limitation is environment dependence. Fingerprinting and channel-charting methods reduce annotation cost, but they remain sensitive to environmental change; the channel-charting paper explicitly notes lifecycle management and updates when the scene changes (Stahlke et al., 2022). RISE likewise depends on human motion and mostly static room structure; completely static environments remain challenging, and fixed-sensor blind regions are not eliminated (Zhou et al., 18 Nov 2025). Passive-reflector systems avoid some of this instability by engineering the environment, but then incur infrastructure planning constraints, visibility constraints, and unknown-reflector-association burdens (Hinderer et al., 19 Sep 2025, Hinderer et al., 19 Sep 2025).

A third limitation is definitional drift. High-accuracy one-node systems such as DragonFly and RISE are often cited as “single radar” references, yet their performance fundamentally depends on MIMO aperture or separate AoA/AoD processing and therefore does not transfer directly to a literal one-channel design (Harisha et al., 7 Jul 2025, Zhou et al., 18 Nov 2025). Conversely, the multi-site SISO literature shows that strict single-channel hardware can remain competitive if multiple nodes are spatially distributed and fused without stringent phase synchronization (Qin et al., 17 Apr 2026).

Current directions point toward three converging lines. One is environment engineering, including optimized passive reflector placement and possibly reconfigurable surfaces; MetaRadar and MF-RIS coexistence work both suggest that controlled propagation can increase location separability, though neither is a direct strict single-channel monostatic solution (Zhang et al., 2020, Tishchenko et al., 6 Feb 2026). A second is map-aware multipath reasoning, in which virtual anchors, ghost inversions, or learned room priors transform NLoS structure into geometric constraints (Leitinger et al., 2014, Zhou et al., 18 Nov 2025). A third is local-window generative optimization, as exemplified by radar-inertial Gaussian-splatting bundle adjustment, which is not a single-channel method but suggests that dense differentiable radar map alignment may eventually be adapted to weaker sensing channels under stronger motion priors (Kung et al., 15 Apr 2026).

Taken together, the literature supports a precise conclusion. A single-channel radar-based indoor localization system is feasible, but only by redistributing complexity: from antenna hardware into environmental references, distributed node geometry, multipath-aware inference, or learned priors. The decisive technical variables are bandwidth, path resolvability, reflector geometry, synchronization assumptions, and the extent to which multipath is modeled as information rather than discarded as interference.