Building Visual Positioning System (VPS) for indoor environments.

Visual Positioning System (VPS) that exists today out of the box is not accurate and reliable in indoor environments because of the following reasons:

• Limited Features: Indoor environments often lack distinct, easily recognizable features such as landmarks, natural textures, or geometric patterns that VPS systems rely on for accurate positioning. Plain walls, minimalistic designs, and uniform surfaces make feature detection and matching challenging.

• Repetitive Environments: Many indoor spaces, like office buildings, shopping malls, or warehouses, have repetitive layouts and structures. Similar-looking corridors, shelves, or rooms create ambiguity for VPS systems, leading to errors in localization and positioning.

• Artificial Lighting: Indoor spaces are illuminated by artificial lights, which can vary in intensity, colour temperature, and flickering patterns. These factors can affect the visibility and appearance of features, making it difficult for VPS systems to maintain consistent and accurate positioning under different lighting conditions.

• Narrow Field of View (FoV): Indoor environments often limit the field of view due to confined spaces, furniture, or obstacles. Narrow angles or blocked perspectives reduce the system's ability to capture and analyze sufficient data for accurate localization, especially in crowded or cluttered environments.

• How is MultiSet optimising VPS to work better in a challenging indoor environment?
  • Spatial Models Optimized for Indoor Environments: We have designed our VPS pipeline specifically to excel in indoor scenes. This includes creating an indoor-focused image retrieval system tailored to handle repetitive and less distinctive features commonly found indoors. Additionally, our pipeline leverages a feature matcher trained specifically for indoor scenes, enabling robust feature detection and matching even in challenging environments. Finally, a pose optimization module has been developed to ensure accurate and stable positioning, compensating for the complexities of indoor layouts.

1. • Comprehensive Indoor Scene Dataset: To enhance the performance of our pipeline, we have built a diverse and challenging dataset representing a wide range of indoor environments. These include locations such as shopping malls, bus stations, train stations, and university campuses. By capturing datasets across varying lighting conditions, architectural styles, and crowd densities, we have fine-tuned each step of the pipeline to better adapt to the unique challenges of indoor environments, ensuring reliable and accurate localization in diverse scenarios.

1. • Advanced Deep Learning Architectures for Indoor Localization: Our system leverages cutting-edge deep learning models that incorporate both global and local features for enhanced scene understanding. To address areas with limited distinctive features, we introduce a feature ranking mechanism that prioritizes the most reliable features for accurate localization. This dual-feature approach ensures that even in feature-sparse environments, the VPS system can maintain high precision and reliability.
   • High-Precision Data Capture via Mapper App: Our proprietary Mapper app enables highly accurate data collection using mobile devices. It synchronizes RGB data, camera matrix, and depth information at a millisecond level, ensuring that all captured data is aligned perfectly in time. This level of precision allows for the creation of high-fidelity 3D maps and datasets, which are critical for accurate indoor localization. By leveraging widely available mobile devices, we make the process scalable and user-friendly while maintaining exceptional data quality.

• In conclusion, while MultiSet's efforts have significantly improved the accuracy and reliability of VPS in challenging indoor environments, we recognize that there is still much work to be done. Building a VPS that performs as seamlessly indoors as it does outdoors remains a complex challenge, requiring continuous innovation in spatial modeling, dataset diversity, deep learning architectures, and data capture methods. Our commitment to advancing these areas ensures that we are constantly refining and optimizing our pipeline to overcome the unique hurdles of indoor environments.