RSSi-Based Visitor Tracking in Museums via Cascaded AI Classifiers and Coloured Graph Representations

E. Onofri, Alessandro CorbettaCollective Dynamics 2022 · DOI

TL;DR

This paper introduces a novel method for accurately tracking museum visitors using noisy RSSi signals from portable beacons, even with a sparse antenna network. By combining AI classifiers with a graph-based representation of museum topology, they achieve over 96% localization accuracy, a significant improvement over previous methods.

What Problem Does This Paper Address?

Understanding visitor behavior in museums is crucial for improving exhibitions, visitor comfort, and artwork preservation. A key aspect of this is tracking individual visitor movements. While portable radio beacons (like Bluetooth or Wi-Fi devices) offer a non-invasive way to do this, the Received Signal Strength Indication (RSSi) they emit is notoriously noisy and fluctuates significantly. This is especially true in historic buildings, where installing a dense network of antennas is often impractical due to architectural constraints and limited power outlets. Existing methods, even those using machine learning like Multi-Layer Perceptrons (MLPs) on short RSSi sequences, struggle to achieve high accuracy in such sparse and complex environments, often failing to account for the museum's inherent structure.

Methods

The core of this paper's contribution lies in a two-pronged approach:

  • Total-Coloured Graph Representation: The museum's layout and its conceptual subdivisions (like exhibition types) are modeled as a "total-coloured graph." Rooms are represented as vertices, and connections (doors, stairs) as edges. Both vertices and edges are assigned "colors" to encode information:
  • Vertex Coloring: Represents typological information (e.g., different exhibition themes).
  • Edge Coloring: Represents architectural constraints (e.g., doors, staircases). This graph structure allows for the definition of room-level and trajectory-level metrics that capture the museum's topology.

  • Cascaded AI Classifiers: Instead of a single complex classifier, the authors propose a "cascaded" system. This system leverages the clustering derived from the coloured graph.

  • Large-Scale Classifier (e.g., Floor Selector): This first layer identifies which broad area or cluster the visitor is in (e.g., which floor of the museum). This reduces the search space for subsequent classifiers.
  • Fine-Scale Classifiers (e.g., Room Selectors): For each identified cluster, a dedicated classifier then pinpoints the visitor's specific room within that cluster. This hierarchical approach allows each classifier to focus on a smaller, more manageable subset of antennas and a simpler problem, leading to improved accuracy and robustness against noisy RSSi signals. They explore various individual localizer choices, including heuristic methods like "argMax" and data-driven methods like MLPs, LSTMs, and Random Forests.

Datasets

The primary dataset used for training and evaluation comes from an experimental campaign conducted at the Galleria Borghese museum in Rome, Italy. This dataset consists of over 900 real-life visitor trajectories, collected using commercial Bluetooth Low-Energy (BLE) beacons and a sparse network of 14 receiving antennas. The paper does not introduce a new benchmark dataset but rather uses this real-world, challenging environment to validate their proposed method.

Key Insights and Ideas

  • Leveraging Museum Topology is Key: The most significant insight is that incorporating explicit knowledge of the museum's structure (rooms, connections, exhibition types) through graph coloring dramatically improves localization accuracy, especially in sparse antenna scenarios.
  • Cascading for Robustness: Breaking down the complex localization problem into a hierarchy of simpler, specialized classifiers (e.g., floor first, then room) is more effective than a single, all-encompassing model. This reduces noise sensitivity and simplifies the learning task for each component.
  • Graph Metrics for Deeper Evaluation: Beyond simple accuracy, the total-coloured graph enables richer evaluation metrics (like mean sample displacement) that provide a more nuanced understanding of trajectory reconstruction quality.
  • RSSi Noise is Manageable with Structure: The paper demonstrates that even with the inherent noisiness and fluctuations of RSSi signals, a structured approach can yield highly accurate results.

Key Results

The proposed cascaded AI classifier method, combined with the total-coloured graph representation, achieved a localization accuracy of over 96.8% in the Galleria Borghese museum. This represents a significant improvement over their previous work, which achieved around 85.8% accuracy using a single MLP. Furthermore, the new method resulted in a much smaller mean displacement error (around 0.08 rooms) compared to previous approaches. The paper highlights that using a symmetric time window for RSSi signal analysis generally yields better results for the fine-scale classifiers.

Limitations and Future Work

  • Antenna Density: While the method is designed for sparse antenna networks, extremely low densities might still pose challenges. The paper mentions that the number of antennas (A=14) was limited by available electrical outlets.
  • Generalizability: The current work focuses on a single museum. Future work could explore automatic methods for identifying optimal features and time scales for each localizer layer, potentially making the approach more adaptable to different museum layouts and conditions.
  • Real-time Performance: While the training is relatively fast, the paper mentions the need to compare online (asymmetric time window) and offline (symmetric time window) performance, suggesting that further optimization for real-time applications might be beneficial.
  • Extending to Multi-Building Museums: The cascaded structure is noted as being extendable to museums with multiple buildings or more complex hierarchical structures, which is a promising avenue for future research.

Summary updated

Abstract

Individual tracking of museum visitors based on portable radio beacons, an asset for behavioural analyses and comfort/performance improvements, is seeing increasing diffusion. Conceptually, this approach enables room-level localisation based on a network of small antennas (thus, without invasive modification of the existent structures). The antennas measure the intensity (RSSi) of self-advertising signals broadcasted by beacons individually assigned to the visitors. The signal intensity provides a proxy for the distance to the antennas and thus indicative positioning. However, RSSi signals are well-known to be noisy, even in ideal conditions (high antenna density, absence of obstacles, absence of crowd, ...). In this contribution, we present a method to perform accurate RSSi-based visitor tracking when the density of antennas is relatively low, e.g. due to technical constraints imposed by historic buildings. We combine an ensemble of "simple" localisers, trained based on ground-truth, with an encoding of the museum topology in terms of a total-coloured graph. This turns the localisation problem into a cascade process, from large to small scales, in space and in time. Our use case is visitors tracking in Galleria Borghese, Rome (Italy), for which our method manages >96% localisation accuracy, significantly improving on our previous work (J. Comput. Sci. 101357, 2021).
BibTeX
@inproceedings{onofri2022rssi,
  author = {E. Onofri and Alessandro Corbetta},
  title = {{RSSi-Based Visitor Tracking in Museums via Cascaded AI Classifiers and Coloured Graph Representations}},
  booktitle = {Collective Dynamics},
  year = {2022},
  doi = {10.17815/cd.2021.131},
}
Figure