Research map / Label-efficient and robust learning / Domain adaptation and generalization
Multitarget Device-Free Localization via Cross-Domain Wi-Fi RSS Training Data and Attentional Prior Fusion
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a Swin-Transformer backbone with a CBAM-based attentional prior fusion module for multitarget device-free localization from Wi-Fi RSS, trained on a mixed real, simulated, and augmented dataset.
Problem
Accurately localizing an unknown number of people with Wi-Fi RSS device-free localization, especially nearby targets with multipath interference and limited real-world multitarget training data.
Approach
Treat multitarget DFL as 2D inverse imaging: a Swin-Transformer maps calibrated RSS to a probability map, and a doubled CBAM prior fusion module fuses a prior localization (e.g., WIMF-RTI output) with the initial estimate; training mixes small real bottle data with simulated and augmented multitarget data.
Key results
- At 40 nodes on T1-5, the method reaches CAcc 0.712 and OSPA 21.301, versus 0.617 and 25.893 for the best baseline (WIMF-RTI).
- With 16 nodes, T1-5 CAcc is 0.234 (OSPA 55.674) and the best baseline WIMF-RTI reaches 0.057 CAcc, so the method is competitive with fewer nodes.
- Training only on simulated data yields T1-5 OSPA 65.41 without prior, while mixing real, simulated, and augmented data gives 33.62 without prior.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed Swin Transformer and prior-fusion method instead of WIMF-RTI | Wi-Fi RSS multitarget localization · T1-5 (M=40) | CAcc: 0.617 -> 0.712 | Helped |
| use the proposed Swin Transformer and prior-fusion method instead of WIMF-RTI | Wi-Fi RSS multitarget localization · T1-5 (M=40) | OSPA: 25.893 -> 21.301 | Helped |
| replace Swin Transformer with ResNet (ablation) | proposed localization method · T1-2 | CAcc: 0.827 -> 0.853 | Hurt |
| replace Swin Transformer with ViT (ablation) | proposed localization method · T1-5 | OSPA: 22.429 -> 21.301 | Hurt |
| use simulation-only training instead of mixed real-simulation-augmentation training (ablation) | proposed localization method · T1-5 | OSPA: 21.30 -> 29.27 | Hurt |
| use real 1-2 target data and simulated 3-5 target data instead of mixed training (ablation) | proposed localization method · T1-5 | OSPA: 21.30 -> 24.69 | Hurt |
| use real 1-2 target data and augmented 3-5 target data instead of mixed training (ablation) | proposed localization method · T1-5 | OSPA: 21.30 -> 21.55 | Hurt |
| replace doubled CBAM with a one-layer convolution of kernel size 1 (ablation) | prior fusion module · T1-5 | CAcc: 0.428 -> 0.712 | Hurt |
| replace doubled CBAM with AFF (ablation) | prior fusion module · T1-5 | OSPA: 27.953 -> 21.301 | Hurt |
| remove the loss constraint on the initial estimation (ablation) | proposed localization method · T1-5 | CAcc: 0.711 -> 0.712 | Hurt |
Surprising findings
- Adding a prior from a previous method does not always help: with fewer nodes (16 or 20) the priors are unreliable, so the authors also evaluate models without the prior fusion module.
Limitations
- Real-world multitarget training data is limited to one or two nearby water bottles; 3-5 target performance relies on simulated and augmented data and shows a larger domain gap.
- Performance drops substantially with fewer Wi-Fi nodes (16 nodes gives CAcc 0.234 on T1-5).
- Post-processing requires a threshold on target count tuned on validation data.
Experimental setup
Datasets: real-world water bottle RSS collections B1, B2-60, B2-42 simulated multitarget dRSS augmented multitarget dRSS P5-42 real-world people test set
Benchmarks: P5-42 real-world test set (T1-2, T3-5, T1-5)
Metrics: CAcc OSPA
Baselines: GVDA-RTI SCLA-RTI WIMF-RTI ResNet ViT
Builds on: Swin Transformer CBAM radio tomographic imaging WIMF-RTI
Closest papers by meaning
- CSI-Bench: A Large-Scale In-the-Wild Dataset for Multi-task WiFi Sensing (NeurIPS 2025 · similarity 0.80)
- ConSense: Continually Sensing Human Activity with WiFi via Growing and Picking (AAAI 2025 · similarity 0.79)
- Person-in-WiFi 3D: End-to-End Multi-Person 3D Pose Estimation with Wi-Fi (CVPR 2024 · similarity 0.79)
- Model-Driven Deep Neural Network for Enhanced AoA Estimation Using 5G gNB (AAAI 2024 · similarity 0.78)
- WiFi CSI Based Temporal Activity Detection via Dual Pyramid Network (AAAI 2025 · similarity 0.78)
- Physics-Informed Generative Modeling of Wireless Channels (ICML 2025 · similarity 0.77)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.