INDOOR VISIBLE LIGHT LOCALIZATION METHOD BASED ON EMBEDDED ARTIFICIAL INTELLIGENCE

Fengwang Zhang, Wenhui Ke, Haoyong Ouyang, Shi Qiu2022 · DOI

Abstract

This paper proposes an indoor visible light location method based on an embedded platform and optical frequency image recognition technology with artificial intelligence, which can effectively improve the location effect in complex indoor environments. By transplanting an artificial intelligence (AI) based image classification algorithm into the embedded platform, this method uses a forward neural network to analyze position information coming from the coded optical frequency image received by a camera. In view of the "motion state" and "occlusion state" that are most likely to fail in traditional visible image localization, we supplement the training set with relevant characteristic optical frequency images to enhance robustness. According to the track and positioning results of the moving platform receiver, the proposed method provides accurate and reliable navigation and positioning with stronger anti-interference ability compared with traditional light intensity or light image positioning methods.

Introduction

Global Positioning Systems (GPS) are the gold standard for outdoor navigation but fail in indoor environments where satellite signals cannot penetrate. While radio-frequency (RF) solutions like Wi-Fi or Bluetooth exist, they often suffer from multi-path interference and signal attenuation in complex environments like mines or shopping malls. Visible Light Positioning (VLP) is a promising alternative that utilizes existing LED infrastructure. However, traditional VLP methods—which usually rely on measuring light intensity or manually decoding "flicker" patterns—are fragile. If a person walks between the light and the sensor (occlusion) or if the sensor moves too quickly (motion blur), the decoding process typically breaks. This research aims to bridge this gap by making VLP robust enough to handle imperfect visual data through embedded AI.

System Architecture and Methodology

The researchers developed a full-stack hardware and software solution designed for real-time indoor localization. The system consists of a transmitter unit where LED lamps are controlled by microprocessors to blink at specific frequencies (e.g., 200Hz, 400Hz, 800Hz, and 1600Hz). These frequencies are high enough to be invisible to the human eye but detectable by a CMOS camera.

System architecture of the visible light localization system, showing the flow from signal generation to neural network analysis.

The core of the receiver is an optical signal acquisition analyzer. Instead of traditional threshold-based decoding, the system treats localization as an image classification problem. A forward neural network, trained using TensorFlow/Keras, is compressed and transplanted onto an ARM Cortex-M4/M7 (STM32) microcontroller using the STMicroelectronics Cube-AI toolset. This allows the "intelligence" to run locally on a small, low-power mobile robot without needing a PC.

The Rolling Shutter Effect and Data Encoding

The system leverages the rolling shutter effect inherent in CMOS cameras (like the OV7670). Because the camera captures the image row-by-row, the high-frequency blinks of the LED appear as distinct horizontal stripes of light and shade.

Example of a high-frequency encoded stripe image captured by the CMOS camera, illustrating the rolling shutter effect.

A 4FSK (Four-level Frequency Shift Keying) coding structure was implemented. A data frame consists of a start bit, a room number, a position number, and a stop bit. By training the neural network to recognize the visual signature of these frequencies even when they are blurry or partially obscured, the system achieves significantly higher resilience than hard-coded decoders.

Experimental Results and Insights

The system was evaluated across several real-world scenarios, focusing on robustness:

  • Static Accuracy: In clear environments, the system achieved a 100% success rate in identifying location beacons, with an average processing time of 45ms.
  • Resilience to Occlusion: Even with partial lateral occlusion, the success rate remained high at 96%. In extreme "corner" cases with significant image defects, the rate only dipped to 88%.
  • Dynamic Performance: On a mobile platform, the system maintained a 90-95% success rate at speeds up to 50cm/s, even on complex routes with 90-degree turns. This was achieved by explicitly including motion-distorted images in the training dataset.
  • Ambient Light Robustness: The system maintained a 100% parsing rate under strong ambient light conditions (e.g., full room illumination and open curtains).

Conclusion

This research demonstrates that treating visible light positioning as an image classification task significantly enhances robustness against common failure modes like motion blur and occlusion. By deploying these models on standard microcontrollers, the authors have proven that high-precision, AI-driven VLP is feasible for low-power mobile robotics. Future work will focus on improving performance at greater heights and higher receiver speeds to accommodate industrial and drone applications.

Summary updated

Abstract

This paper proposes an indoor visible light location method based on embedded platform and optical frequency image recognition technology with artificial intelligence, which can effectively improve the location effect in complex indoor environment. By transplanting the artificial intelligence (AI) based image classification algorithm into the embedded platform, this method uses a forward neural network to analyse the position information coming from the coded optical frequency image received by a camera, and then the positioning results can be obtained. In view of the "motion state" and "occlusion state" that are most likely to fail in traditional visible image localization, we specially supplement the training set of relevant characteristic optical frequency images to enhance the robustness of the method. According to the track and positioning results of the moving platform receiver, the proposed method can provide accurate and reliable navigation and positioning and has stronger anti-interference ability compared with the traditional light intensity or light image positioning methods.
BibTeX
@inproceedings{zhang2022indoor,
  author = {Fengwang Zhang and Wenhui Ke and Haoyong Ouyang and Shi Qiu},
  title = {{INDOOR VISIBLE LIGHT LOCALIZATION METHOD BASED ON EMBEDDED ARTIFICIAL INTELLIGENCE}},
  booktitle = {˜The},
  year = {2022},
  doi = {10.5194/isprs-archives-xlvi-3-w1-2022-255-2022},
}
Figure