Source-linked AI summary
Automotive mmWave Spinning Radar Place Recognition with Spatially Gated Feature-Correlation Representation
Saimunur Rahman, Sagun Singh Shrestha, Abdelwahed Khamis, Peyman Moghadam
TL;DR
Place recognition with spinning radar must handle heading-induced circular shifts and variable, sometimes ambiguous radar returns, while conventional aggregation can lose relationships among responses. SGCA-Net combines rotation-robust feature extraction with spatially gated pairwise correlations; it improves across MulRan settings and generalizes to unseen HeRCULES environments and sensors without fine-tuning.
Problem
Spinning-radar place recognition must handle heading-induced azimuth shifts and variable returns, while conventional aggregation can lose feature relationships that distinguish similar places.
Source: §1 · pp. 1–2Method
SGCA-Net combines a rotation-robust radar backbone with learned spatial keep/dustbin weights and pairwise feature-correlation aggregation.
Source: §1 · p. 2Results
Takeaways & Limitations
The authors report a more discriminative global representation than conventional first-order aggregation.
Source: §6 · p. 7The stated positive-semidefinite property of Σ depends on the spatial weights summing to one.
Source: §3.2 · p. 4
Abstract
from arXiv · showhide
Automotive spinning FMCW radar provides dense, $360^\circ$ sensing and remains reliable under poor illumination and adverse weather, making it well-suited to autonomous navigation. Place recognition uses these observations to identify previously visited locations for re-localization and long-term navigation. However, heading changes appear as circular shifts in the polar radar representation, and conventional global aggregation can lose relationships among radar responses that are important for distinguishing similar places. We propose SGCA-Net, a spinning radar place recognition framework that combines rotation-robust feature extraction with Spatially Gated Correlation Aggregation (SGCA). SGCA learns spatial weights to reduce the influence of unstable and ambiguous radar regions, while aggregating pairwise correlations among local responses to preserve informative feature relationships. Experiments on the MulRan dataset show that SGCA-Net consistently outperforms SOTA methods across urban, campus, and open-road environments, while remaining robust to substantial heading variation. Evaluation on the HeRCULES dataset further demonstrates that SGCA-Net generalizes to unseen environments and radar sensors without fine-tuning.
1. Introduction
Spinning FMCW radar offers dense 360° azimuth-range sensing, but heading changes shift scans and variable returns make place descriptors difficult to keep both repeatable and distinctive. SGCA-Net combines rotation-robust feature extraction with spatially gated pairwise feature correlations, and the paper reports improvements across MulRan environments and transfer to unseen environments and sensors.
1. Introduction
- Spinning FMCW radar provides dense 360° azimuth-range measurements and remains robust under poor illumination and adverse weather.The passage contrasts its dense coverage with the sparser, more limited field of view typical of 4D automotive radar.Source: §1 · p. 1
- Heading changes appear as circular shifts along azimuth, while radar returns vary with scene and sensing factors, so revisits can differ and distinct places can look similar.The cited sources of variation include reflector geometry, material, incidence angle, multipath, sidelobes, range, dynamic objects, occlusions, and transient clutter.Source: §1 · p. 1
- Radar place-recognition work addresses heading variation, but aggregation of local features into global descriptors remains comparatively less explored.Conventional first-order and learned aggregation can lose co-occurrence structure among responses that may help distinguish places.Source: §1 · pp. 1–2
- SGCA-Net combines a rotation-robust radar backbone with spatially gated pairwise feature correlations, reducing unstable or ambiguous locations while preserving repeatable radar structure.The module learns keep/dustbin weights and applies power normalization to stabilize the representation.Source: §1 · p. 2
- SGCA-Net consistently improves over representative radar place-recognition methods across diverse MulRan environments and substantial heading variation, and transfers to unseen environments and sensors without fine-tuning.The transfer result is reported on HeRCULES.Source: §1 · p. 2
2. Related Work
Radar place-recognition methods have progressed from hand-crafted scan representations to learned descriptors, while global aggregation still often emphasizes response strength or distribution over feature interactions. SGCA-Net applies learned spatial gating before pairwise correlation aggregation to reduce the effect of unreliable locations.
2. Related Work
- Learned spinning-radar descriptors include RadarLoc, Open-RadVLAD, RaPlace, and SHeRLoc, while most still use conventional global aggregation.Earlier ScanContext uses circular alignment and Ring Key provides a compact rotation-invariant summary; hand-crafted methods have limited adaptability to radar appearance variation.Source: §2 · p. 2
- Average, max, and GeM pooling summarize individual responses, while NetVLAD and Open-RadVLAD model local-feature distributions without explicitly capturing response relationships.Radar feature maps can contain empty regions, multipath, transient objects, and repeated structures that introduce unstable or ambiguous responses.Source: §2 · p. 2
- Bilinear pooling and covariance representations model feature-channel dependencies, but unreliable radar locations can distort correlation statistics.SGCA-Net gates locations before correlation aggregation and uses matrix square-root normalization to stabilize the representation.Source: §2 · p. 2
- Figure 1 traces the pipeline from a polar BEV radar image through a stem and residual backbone to SGCA processing and a global descriptor.The SGCA block performs residual downsampling, keep-and-dustbin selection, feature-correlation aggregation, power normalization, and vectorization.Source: Figure 1 · p. 3
3. Methodology
SGCA-Net maps polar radar scans through a rotation-robust backbone and SGCA pooling to a compact global descriptor, trained end-to-end with a rotation-robust metric-learning objective.
3. Methodology
- The descriptor pipeline applies a rotation-robust backbone to polar radar input, then uses SGCA to produce the compact global representation.The network is trained end-to-end with a rotation-robust metric-learning objective.Source: §3 · p. 3
3.1. Radar Polar Image Backbone
- Cylindrical azimuth padding preserves boundary continuity for heading-shift robustness; a fusion block combines deep and intermediate features before SGCA.Source: §3.1 · pp. 3–4
3.2. Spatially Gated Correlation Aggregation
- A keep/dustbin gate normalizes spatial keep weights before weighted correlation pooling, emphasizing informative regions while limiting unstable responses and retaining pairwise channel co-occurrence.Source: §3.2 · p. 4
- The symmetric correlation matrix is trace-normalized, square-root normalized with Newton–Schulz iteration, and vectorized into a 2,080-dimensional descriptor.Source: §3.2 · p. 4
3.3. Rotation-Robust Triplet Loss
- Training circularly shifts scans by uniformly sampled azimuth rotations up to 180°, then uses batch-hard positives and negatives in a triplet objective with margin 0.2.Source: §3.3 · p. 4
4. Experimental Setup
The evaluation uses spinning FMCW radar scans from MulRan across geographically disjoint urban, campus, and open-road regions. Place recognition is measured by Recall@1 at specified distance thresholds under a shared training and evaluation protocol.
Dataset and Evaluation Environments
- MulRan scans come from a Navtech CIR204-H FMCW radar with 360° coverage, 0.9° angular resolution, and 0.06 m radial resolution, then are represented as polar images and downsampled to 384 × 128 pixels.The original polar image uses 400 azimuth bins and 3,360 range bins.Source: §4 · pp. 4–5
Evaluation Metric
- Table 1 reports Recall@1% radar place-recognition results on the MulRan evaluation sequences.The supplied caption identifies the table's scope but does not provide result values.Source: Table 1 · p. 5
Training and Evaluation Protocol
- Evaluation spans Sejong, KAIST, and Riverside, representing urban, campus, and open-road conditions, with geographically disjoint regions for training and evaluation.Training uses 13,611 scans from Sejong traversals 01 and 02; traversal 02 supplies the reference database and traversal 01 the query sequence.Source: §4 · p. 5
Evaluation Metric
- Recall@1 ranks database scans by Euclidean distance between global descriptors and counts a retrieval as correct when the top match is within the ground-truth distance threshold.Results are reported at 3 m, 5 m, and 10 m for each environment and as an unweighted mean.Source: §4 · p. 5
5. Results
SGCA-Net improves retrieval over RadarLoc across MulRan thresholds and environments, retains performance under large heading changes, and transfers to HeRCULES without fine-tuning. Ablations and qualitative examples highlight the contribution of correlation aggregation and spatial gating.
5.1. Comparison with Existing Methods
- At 5 m, mean R@1 rises from 87.7% to 91.3%, and at 10 m from 96.6% to 98.1% against RadarLoc; at 3 m it rises from 66.6% to 71.4%.SGCA-Net achieves the highest R@1 on every sequence at both 5 m and 10 m.Source: Figure 2 · p. 5, §5.1 · pp. 5–6
5.2. Ablation Analysis
- Ungated correlation raises mean R@1 from 66.6% to 69.4% at 3 m; full SGCA-Net reaches 71.4%, 91.3%, and 98.1% at 3 m, 5 m, and 10 m.The ablation compares variants with the same training data, optimization settings, and evaluation protocol.Source: Table 2 · p. 6, §5.2 · p. 6
- Under random query rotations up to 180°, SGCA-Net improves mean R@1 over RadarLoc from 65.9% to 70.6% at 3 m, 87.4% to 91.1% at 5 m, and 96.5% to 98.1% at 10 m.Source: Table 3 · p. 6, §5.2 · p. 6
5.3. Transfer to Unseen Radar Environments
- On HeRCULES, SGCA-Net beats RadarLoc on all four session pairs and raises AR@1 from 0.940 to 0.958 after MulRan-only training without fine-tuning.The evaluation includes clear-to-night/cloudy and clear-to-snowy session changes.Source: Table 4 · p. 6, §5.3 · pp. 6–7
5.4. Qualitative Analysis
- In three Sejong examples, RadarLoc returns incorrect top-ranked matches while SGCA-Net correctly recognizes all three places under the 3 m criterion.The examples include broad curved returns, asymmetric scene structure, and open roads with few distinctive landmarks.Source: Figure 3 · p. 7, §5.4 · p. 7
6. Conclusion and Future Work
SGCA-Net combines rotation-robust feature extraction with spatially gated feature-correlation aggregation, producing a more discriminative global representation than conventional first-order aggregation. It improves over representative RPR methods across diverse MulRan environments and heading variation, and generalizes to unseen HeRCULES environments and radar sensors without fine-tuning.
6. Conclusion and Future Work
- SGCA-Net models pairwise radar-feature relationships while reducing the influence of unstable and ambiguous regions.The framework combines rotation-robust feature extraction with spatially gated feature-correlation aggregation.Source: §6 · p. 7
- SGCA-Net provides a more discriminative global representation than conventional first-order aggregation.Source: §6 · p. 7
- SGCA-Net consistently improves over representative RPR methods across diverse environments and substantial heading variation on MulRan.Source: §6 · p. 7
- On HeRCULES, SGCA-Net generalizes to unseen environments and radar sensors without fine-tuning.Source: §6 · p. 7
- Future work will target computational efficiency, broader evaluation, long-term environmental changes, and uncertainty-aware spatial weighting.Source: §6 · p. 7