Abstract: This paper considers indoor localization using multi-modal wireless signals including Wi-Fi, inertial measurement unit (IMU), and ultra-wideband (UWB). By formulating the localization as a multi-modal sequence regression problem, a multi-stream recurrent fusion method is proposed to combine the current hidden state of each modality in the context of recurrent neural networks while accounting for the modality uncertainty which is directly learned from its own immediate past states. The proposed method was evaluated on the large-scale SPAWC2021 multi-modal localization dataset and compared with a wide range of baseline methods including the trilateration method, traditional fingerprinting methods, and convolution network-based methods.
Introduction
Indoor localization remains a significant challenge in "GPS-less" environments where signal blockage and interference are prevalent. Traditional methods often rely on single modalities like Wi-Fi, which is ubiquitous but noisy; Ultra-Wideband (UWB), which is precise but sensitive to line-of-sight (LoS) obstructions; or Inertial Measurement Units (IMUs), which provide good short-term tracking but suffer from long-term drift. The core difficulty lies in managing modality uncertainty—knowing which sensor to trust when environmental conditions change. This work proposes a multi-stream recurrent fusion network that dynamically learns sensor reliability from temporal data history.
Methodology
The proposed architecture, the Multi-Stream Recurrent Fusion Network, employs a dedicated Long Short-Term Memory (LSTM) pipeline for each sensor modality: Wi-Fi Received Signal Strength Indicator (RSSI), Wi-Fi Channel State Information (CSI), UWB, and IMU. This structure enables the model to capture the unique temporal evolution of each signal type.
A key innovation is the Recurrent Fusion Block. Instead of simple concatenation, the model evaluates the immediate past hidden states of each stream to calculate an importance weight ($\alpha$) for each modality. These weights are normalized via a Softmax layer, allowing the system to dynamically prioritize reliable sensors. For instance, if UWB signals are degraded by obstacles, the model automatically shifts weight to IMU or Wi-Fi data. Additionally, raw Wi-Fi CSI is compressed using polynomial fitting to reduce computational overhead while preserving essential signal features.
Experimental Evaluation
The model was evaluated using the SPAWC2021 multi-modal localization dataset, comprising approximately 750,000 samples collected by a robot over a 2-kilometer path. The setup included 11 Wi-Fi anchors and 3 UWB anchors, with high-precision ground truth provided by Lidar-based SLAM.
Results and Discussion
The Recurrent Fusion (RF) method demonstrated substantial improvements over baseline approaches. The optimal configuration—a bi-directional LSTM with a sequence length ($T$) of 10 and a past-state lookback ($F$) of 1—achieved a mean localization error of just 6 cm.
As shown in the performance plots, the fusion method significantly outperforms individual sensors and non-recurrent fusion baselines. For example, the best non-recurrent method (RIU-Conv) had a mean error of 26 cm, making the RF method nearly four times more accurate. The results also indicated that shorter sequence lengths ($T=10$) were more effective for high-speed localization, suggesting that recent data is more critical than long-term history in dynamic indoor environments. Even in "dead zones" where UWB coverage was zero, the fusion model maintained high accuracy by dynamically shifting weight to Wi-Fi and IMU data.
Conclusion
Multi-modal recurrent fusion provides a robust solution for high-precision indoor localization by effectively managing sensor uncertainty. While the current model achieves centimeter-level accuracy, future work could focus on reducing the computational load for deployment on low-power IoT devices and adapting the system to environments with unknown anchor positions.
Summary updated