Hierarchical Stage-Wise Training of Linked Deep Neural Networks for Multi-Building and Multi-Floor Indoor Localization Based on Wi-Fi RSSI Fingerprinting

Sihao Li, Kyeong Soo Kim, Zhe Tang, Graduate, Jeremy S. SmitharXiv 2024 · DOI · arXiv · OpenReview

TL;DR

This paper introduces a novel "Hierarchical Stage-Wise Training" (HST) framework for Wi-Fi RSSI fingerprinting-based indoor localization in large, multi-building, and multi-floor environments. By training linked deep neural networks in stages, leveraging prior knowledge from higher-level tasks to inform lower-level ones, they achieve state-of-the-art 3D localization accuracy on a challenging public dataset.

What Problem Does This Paper Address?

Accurately pinpointing locations indoors is crucial for many applications, from asset tracking to emergency response. While GPS works outdoors, it's unreliable inside due to signal obstruction. Wi-Fi fingerprinting, which uses signal strength patterns to identify locations, is a practical alternative. However, scaling this to massive environments like shopping malls or university campuses presents significant challenges. Existing deep learning approaches often struggle with this scalability, requiring separate training for each building or floor, leading to complex database management and training procedures. This paper tackles the problem of efficiently and accurately localizing devices in such large-scale, hierarchical indoor spaces.

Methods

The core innovation is the Hierarchical Stage-Wise Training (HST) framework. This method extends the concept of stage-wise training (where a model learns progressively) to a system of multiple linked neural networks. Each network is responsible for a specific sub-problem within the localization hierarchy (e.g., identifying the building, then the floor, then the precise location within the floor).

Here's how it works:

  • Hierarchical Structure: The framework explicitly models the inherent hierarchy of indoor localization: Building -> Floor -> Location within Floor.
  • Linked Networks: Instead of a single monolithic network, the system uses multiple interconnected networks, each specialized for a stage in the hierarchy.
  • Stage-Wise Training: Networks are trained sequentially. For instance, the network responsible for floor-level localization is trained after the building and floor identification networks have been trained. Crucially, the knowledge learned by the higher-level networks (building/floor ID) is used to initialize and guide the training of the lower-level network (precise location). This "prior knowledge transfer" is the key to improved performance and scalability.
  • Flexibility: The HST framework is designed to be architecture-agnostic, meaning it can be applied to different types of deep neural networks, such as Feedforward Neural Networks (FNNs) and Convolutional Neural Networks (CNNs). The paper demonstrates this by adapting existing hybrid SIMO DNN and CNNLoc architectures into linked network versions for HST.

Datasets

The primary dataset used for training and evaluation is the UJIIndoorLoc database. This is a publicly available, large-scale dataset specifically designed for multi-building and multi-floor indoor localization challenges. It covers three buildings with four or five floors each on the Universitat Jaume I (UJI) campus in Spain. The database contains Wi-Fi Received Signal Strength Indicator (RSSI) measurements from numerous access points, along with corresponding location information (building, floor, and 3D coordinates).

Key Insights and Ideas

  • Exploiting Hierarchy is Key: The paper strongly emphasizes that simply applying standard deep learning to large-scale indoor localization is suboptimal. The inherent hierarchical nature of the problem (building, then floor, then location) must be explicitly leveraged for efficient and accurate solutions.
  • Knowledge Transfer Across Stages: The HST framework's success hinges on the idea that learning to identify a building and floor provides valuable contextual information that significantly aids in pinpointing the exact location within that floor. This is more effective than trying to learn all aspects simultaneously or in isolation.
  • Scalability Through Modularization: By breaking down the problem into linked sub-networks trained in stages, the approach becomes more manageable and scalable for vast indoor environments, avoiding the need for massive, monolithic models or complex database management.
  • Architecture Agnosticism: The HST framework's ability to work with both FNN and CNN architectures demonstrates its versatility and broad applicability to different deep learning approaches for this problem.

Key Results

The proposed Hierarchical Stage-Wise Training (HST) framework, when applied to linked deep neural networks, achieved state-of-the-art results on the UJIIndoorLoc dataset:

  • Linked-DNN Model: Achieved a 3D localization error of 8.19 m. The authors claim this is the most accurate result ever obtained for neural network-based models on the full UJIIndoorLoc dataset.
  • Linked-CNNLoc Model: Significantly reduced the 3D localization error from 11.78 m (original CNNLoc) to 8.71 m when trained with the HST framework.
  • Improved Accuracy: In general, the proposed HST models (both DNN and CNN variants) consistently outperformed their conventionally trained counterparts in terms of average 3D error.
  • Robustness: While floor hit rates saw a slight decrease for the CNN model, overall building hit rates remained at 100%, and the 3D error improvements were substantial.
  • Training Efficiency: Despite training multiple linked networks in stages, the proposed models showed significantly lower training times compared to hierarchical RNN models, making them more practical for large-scale deployments.

Limitations and Future Work

  • Floor Hit Rate: The paper notes a slight decrease in the floor hit rate for the proposed CNN model compared to some other methods.
  • Generalizability to Other Sensors: While the paper suggests the HST framework can be extended to other wireless signal sensors (BLE, Zigbee, LoRaWAN), this is presented as future work.
  • Data Augmentation: The paper mentions that some state-of-the-art models (like CNNLoc) rely on complex data augmentation, whereas their proposed HST approach achieves competitive results without it, implying potential for even further improvement with augmentation.
  • Computational Advantage: While faster than RNNs, the training time for HST models is slightly higher than conventionally trained single models, a trade-off for the improved accuracy and hierarchical exploitation.

Future work could focus on further optimizing the stage-wise training process, exploring its application to other sensor modalities, and investigating advanced data augmentation techniques within the HST framework to push localization accuracy even further.

Summary updated

Abstract

In this paper, we present a new solution to the problem of large-scale multi-building and multi-floor indoor localization based on linked neural networks, where each neural network is dedicated to a sub-problem and trained under a hierarchical stage-wise training framework. When the measured data from sensors have a hierarchical representation as in multi-building and multi-floor indoor localization, it is important to exploit the hierarchical nature in data processing to provide a scalable solution. In this regard, the hierarchical stage-wise training framework extends the original stage-wise training framework to the case of multiple linked networks by training a lower-hierarchy network based on the prior knowledge gained from the training of higher-hierarchy networks. The experimental results with the publicly-available UJIIndoorLoc multi-building and multi-floor Wi-Fi RSSI fingerprint database demonstrate that the linked neural networks trained under the proposed hierarchical stage-wise training framework can achieve a three-dimensional localization error of 8.19 m, which, to the best of the authors' knowledge, is the most accurate result ever obtained for neural network-based models trained and evaluated with the full datasets of the UJIIndoorLoc database, and that, when applied to a model based on hierarchical convolutional neural networks, the proposed training framework can also significantly reduce the three-dimensional localization error from 11.78 m to 8.71 m.
BibTeX
@misc{li2024hierarchical,
  author = {Sihao Li and Kyeong Soo Kim and Zhe Tang and Graduate and Jeremy S. Smith},
  title = {{Hierarchical Stage-Wise Training of Linked Deep Neural Networks for Multi-Building and Multi-Floor Indoor Localization Based on Wi-Fi RSSI Fingerprinting}},
  year = {2024},
  eprint = {2407.13288},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2407.13288},
}

Loading related content...

Figure