Abstract
Existing indoor localization methods face significant constraints, including multipath effects in Wi-Fi, high costs for ultra-wide-band (UWB), and poor anti-interference for Bluetooth. To address these, we propose a vision-based indoor localization method. The environment is divided into regions, each with a location center. In offline mode, a VGG16 network is pre-trained on ImageNet and fine-tuned on a custom indoor dataset. In online mode, the converged network processes a mobile robot's RGB video stream to extract location-specific features, which are then classified by an ArcFace classifier. Experimental results demonstrate that this method accurately estimates locations in cluttered, unstructured scenes with 94.7% accuracy without additional hardware.
Introduction
Indoor localization is a notoriously difficult problem. While GPS is the standard for outdoor positioning, indoor environments obstruct satellite signals. Traditional workarounds like Wi-Fi fingerprinting or Bluetooth beacons are prone to signal bouncing (multipath effects) and interference. High-precision alternatives like UWB are often cost-prohibitive for large-scale deployment. Furthermore, traditional computer vision approaches often struggle with "visual aliasing," where different parts of a building look nearly identical.
This research introduces a low-cost, robust system requiring only a standard RGB camera. By treating localization as a feature-matching problem similar to facial recognition, the system achieves high reliability in complex industrial environments.
Methodology
The proposed pipeline utilizes transfer learning and advanced classification to turn visual data into spatial coordinates.
Feature Extraction and Transfer Learning
Rather than training a model from scratch, the system uses a VGG16 architecture pre-trained on ImageNet. The first 13 convolutional layers are fine-tuned using a custom dataset of indoor images. This allows the model to retain general visual recognition capabilities while specializing in the specific architectural features of the target environment.
The ArcFace Classifier
A key innovation of this method is the replacement of the standard Softmax classifier with ArcFace (Additive Angular Margin Loss). While Softmax is effective for distinct object classes, it can struggle with the subtle differences between similar-looking rooms. ArcFace forces the network to learn more "compact" features for the same location and creates a wider angular margin between different locations, significantly improving discriminative power.
Experimental Setup and Results
The system was tested in an industrial warehouse environment divided into nine distinct regions. The custom dataset included 100 training images and 21 testing images per location, captured at various angles to ensure orientation robustness.
- High Accuracy: The system achieved a peak localization accuracy of 94.7%.
- Classifier Performance: The ArcFace method outperformed the standard Softmax classifier by 3.7%.
- Efficiency: Through transfer learning, the model reached convergence in just 35 epochs with a training loss of 0.03.
Conclusion
This vision-based approach provides a high-accuracy, infrastructure-free solution for indoor robot navigation. By leveraging the discriminative power of ArcFace, the system effectively handles visual aliasing in cluttered scenes. Future work will focus on improving robustness against dynamic obstacles (such as humans blocking the camera view) and scaling the system to larger, more complex industrial complexes.
Summary updated