Indoor localization using light spectral information

Yanxiang Wang, Jiawei Hu, Hong Jia, Wen Hu, Mahbub Hassan, Ashraf Uddin, Brano Kusy, Moustafa YoussefMobiCom 2022 · OpenReview

Abstract

Indoor localization remains a significant challenge in ubiquitous computing, as traditional GPS signals are unavailable and radio-based methods (WiFi/Bluetooth) are often susceptible to environmental noise. This paper presents Spectral-Loc, a passive indoor localization system that utilizes the spectral distribution of ambient light—its "color"—rather than mere intensity. By leveraging the unique reflectance properties of indoor materials across 18 wavelength sub-bands, Spectral-Loc achieves sub-meter accuracy using commercial-off-the-shelf (COTS) sensors. Our evaluation in a complex office environment demonstrates a 90th percentile error of 0.98 meters, significantly outperforming traditional intensity-based approaches.

1. Introduction

Accurate indoor localization is essential for applications ranging from warehouse logistics to assistive technologies. While existing solutions like Visible Light Communication (VLC) offer high precision, they typically require "modulated light," necessitating expensive retrofitting of existing lighting infrastructure with high-frequency flickering LEDs. Conversely, passive light-based systems that rely on intensity (brightness) are often ineffective; multiple locations within a room frequently share identical light levels, leading to high localization ambiguity.

Spectral-Loc addresses these limitations by shifting the focus from light intensity to spectral composition. By analyzing how light interacts with the specific materials in an environment—such as furniture, paint, and flooring—the system creates a unique "spectral fingerprint" for every location.

2. System Design

2.1 Hardware and Sensing

The system is prototyped using the AS7265x, a low-cost, COTS spectral sensor capable of measuring light across 18 discrete wavelength sub-bands spanning from the ultraviolet (UV) to the near-infrared (NIR) spectrum. This multi-band approach allows the system to capture subtle differences in the light environment that are invisible to standard photodiodes.

2.2 Spectral Fingerprinting and Normalization

The core innovation of Spectral-Loc lies in its use of reflectance. Different materials (e.g., a metal cabinet vs. a wooden bookshelf) reflect light differently across the spectrum. Even if two spots receive identical illumination from a ceiling fixture, the light reflected from the surrounding environment to a user's sensor will carry a unique spectral signature based on the proximity and composition of nearby objects.

To ensure robustness against changes in overall illumination—such as a bulb dimming or a secondary lamp being activated—the system employs a normalization technique. This process focuses on the relative distribution of power across the 18 bands rather than the absolute power, ensuring the fingerprint remains stable despite fluctuations in total brightness.

2.3 Deep Learning Model

To map these spectral fingerprints to physical coordinates $(x, y)$, we designed a Convolutional Neural Network (CNN) enhanced with an attention mechanism. The CNN extracts spatial features from the 18-band spectral data, while the attention mechanism helps the model prioritize the most informative wavelength bands for a given environment.

3. Experimental Evaluation

3.1 Datasets and Environments

Spectral-Loc was evaluated in two distinct settings:

  1. Meeting Room: A controlled 7.4m x 3.9m space.
  2. Large Office: A complex 13.3m x 8.3m environment featuring cubicles, diverse furniture materials (metal, wood, plastic), and varied carpet textures.

Data was collected over several days under varying lighting configurations to test the system's resilience to environmental changes.

3.2 Performance Results

In the large office environment, Spectral-Loc achieved a 90th percentile error of 0.98 meters. In comparison, a traditional intensity-based system under the same conditions yielded an error exceeding 5 meters.

Furthermore, the system demonstrated remarkable robustness to hardware failure. In a simulated scenario where an entire row of ceiling lights was deactivated, the normalized Spectral-Loc maintained a median error of 0.31m. Under the same failure conditions, the intensity-based method's error jumped to 3.25m, representing a tenfold difference in reliability.

4. Key Insights

  • Spectral Stability: While light intensity fluctuates throughout the day, the relative spectral distribution at a specific location remains remarkably stable, providing a more reliable anchor for localization.
  • Band Optimization: Analysis revealed that localization accuracy plateaus at approximately 8 sub-bands. This suggests that future iterations could utilize even simpler, more energy-efficient sensors without sacrificing significant precision.
  • Sensor Placement: Experimental data indicated that sensors worn on the upper body (chest or back) outperform lower-body placements. This is likely due to a clearer "view" of the reflected environment and fewer occlusions from the user's own movements.

5. Limitations and Future Work

Despite its high accuracy, Spectral-Loc currently faces several hurdles for wide-scale deployment. The presence of sunlight introduces a dynamic spectral bias that can overwhelm the subtle reflections of indoor materials; current testing has been limited to nighttime or windowless environments. Additionally, the model's sensitivity to user orientation and the transition from static spots to continuous "walking traces" remain areas for active research. Future work will also focus on reducing the power consumption of spectral sensors to match the efficiency of simple photodiodes.

6. Conclusion

Spectral-Loc demonstrates that the spectral composition of ambient light is a powerful, underutilized resource for indoor localization. By leveraging the unique reflectance of indoor environments and employing deep learning for feature extraction, the system provides a high-precision, infrastructure-free solution that significantly outperforms traditional light-intensity methods.

Summary updated

BibTeX
@inproceedings{wang2022indoor,
  author = {Yanxiang Wang and Jiawei Hu and Hong Jia and Wen Hu and Mahbub Hassan and Ashraf Uddin and Brano Kusy and Moustafa Youssef},
  title = {{Indoor localization using light spectral information}},
  booktitle = {MobiCom},
  year = {2022},
  url = {https://openreview.net/forum?id=oYvFfcrdqn},
}

Loading related content...

Figure