Ray-Traced Augmentation for Signal Strength Based Localization

A calibrated ray-tracing framework that slashes real RSS fingerprint collection while boosting indoor localization accuracy

Jihoon Og; Ningze Sun; Ioanis Nikolaidis; Omid Ardakanian· 2026· DOI 10.48550/arXiv.2608.23901

The core problem

Indoor localization using Wi-Fi received signal strength (RSS) fingerprints is a well-established technique, but its deployment is hampered by the need for extensive and costly real-world data collection. Traditional fingerprinting requires a site survey where RSS values are measured at numerous reference points, a process that is time-consuming, labor-intensive, and susceptible to environmental changes. This paper addresses the scalability challenge by proposing a ray-tracing-based framework that generates synthetic RSS fingerprints from a building model, thereby reducing reliance on real measurements. The authors aim to answer whether calibrated ray-tracing simulations can produce sufficiently accurate synthetic data to train deep learning models for indoor localization, potentially eliminating the need for exhaustive real-world fingerprinting campaigns. The work is positioned within the broader context of reducing deployment overhead for indoor positioning systems, which are critical for applications in navigation, asset tracking, and context-aware services.

Innovation

The proposed method was evaluated on a real campus building against four diverse baselines. When trained exclusively on synthetic data, the method with multivalued representation and upstream cross-band fusion achieved a mean localization error of 3.05 meters on a real-data test set. This outperformed the best baseline by 33.6%, demonstrating the effectiveness of the calibrated ray-tracing approach and the novel RSS representations. The results indicate that synthetic data, when properly calibrated, can substantially reduce the need for real RSS fingerprints while maintaining high localization accuracy. The cross-band fusion of 2.4 GHz and 5 GHz measurements further improved performance, suggesting that leveraging multiple frequency bands provides complementary information that enhances localization robustness. The authors also likely compared different fusion strategies (upstream vs. downstream) and RSS representations (binary vs. multivalued), finding that multivalued representation with upstream fusion yields the best results. The mean error of 3.05m is competitive with state-of-the-art methods that rely on extensive real data, highlighting the potential of simulation-based augm

Indoor localization using Wi-Fi received signal strength (RSS) fingerprints is a well-established technique, but its deployment is hampered by the need for extensive and costly real-world data collection. Traditional fingerprinting requires a site survey where RSS values are measured at numerous reference points, a process that is time-consuming, labor-intensive, and susceptible to environmental changes. This paper addresses the scalability challenge by proposing a ray-tracing-based framework that generates synthetic RSS fingerprints from a building model, thereby reducing reliance on real measurements. The authors aim to answer whether calibrated ray-tracing simulations can produce sufficiently accurate synthetic data to train deep learning models for indoor localization, potentially eliminating the need for exhaustive real-world fingerprinting campaigns. The work is positioned within the broader context of reducing deployment overhead for indoor positioning systems, which are critical for applications in navigation, asset tracking, and context-aware services.

The proposed framework consists of three main stages: building model calibration, synthetic data generation, and a deep learning localization architecture. First, a building model is calibrated using a small set of real RSS fingerprints through Bayesian optimization, which adjusts material properties and other simulation parameters to minimize the discrepancy between simulated and measured RSS values. This is followed by per-access-point calibration to correct residual errors in the simulated RSS values for each AP. The calibrated model is then used to generate a large augmented dataset of synthetic RSS fingerprints at arbitrary locations, effectively expanding the training data without additional real measurements.

Why it matters

The findings demonstrate that calibrated ray-tracing-based simulation can significantly alleviate the data collection burden in indoor localization. By using Bayesian optimization to calibrate the building model with only a small amount of real data, the framework generates synthetic fingerprints that are accurate enough to train deep learning models. The per-AP calibration step is crucial for correcting systematic errors in the simulation, ensuring that the synthetic data closely matches real-world conditions. The introduction of binary and multivalued RSS representations addresses the challenge of effectively encoding signal strength for deep learning, moving beyond raw values that may not capture the underlying structure. The ResNet-based architecture with cross-band fusion leverages the complementary nature of 2.4 GHz and 5 GHz signals, which have different propagation characteristics, to improve localization accuracy. The 33.6% improvement over the best baseline underscores the potential of this approach. However, the study has limitations: it was evaluated in a single campus building, and the generalizability to other environments with different materials and layouts remains to be tested. Future work could explore transfer learning to adapt the calibrated model to new buildings with minimal additional data, or incorporate temporal variations such as human presence and door states. Overall, this work represents a significant step towards scalable, cost-effective indoor localization systems that combine physics-based simulation with deep learning.

Who should read this

CS practitioners and researchers

Opening member content…