A Sheaf Theoretical Approach to Uncertainty Quantification of Heterogeneous Geolocation Information

Cliff A. Joslyn, Lauren E. Charles, Chris DePerno, Nicholas Gould, Kathleen Nowak, Brenda Praggastis, Emilie Purvine, Michael Robinson, Jennifer Strules, Paul WhitneySensors 2020 · OpenReview

TL;DR

This paper introduces a novel approach to integrating data from heterogeneous sensors for tracking applications, using the mathematical framework of sheaves. This method provides a theoretically grounded way to quantify sensor agreement and diagnose data inconsistencies, outperforming traditional methods when sensor data isn't perfectly aligned.

What Problem Does This Paper Address?

Tracking multiple targets with diverse sensors (e.g., GPS, radio telemetry, text reports) is a significant challenge. Each sensor has its own error model, and combining their data meaningfully is difficult. Existing methods often struggle when sensor data isn't in a common coordinate system or when sample rates are irregular. The authors aim to develop a unified, theoretically justified framework for quantifying the agreement among these heterogeneous sensors, both globally and at a fine-grained level, and to diagnose issues when data doesn't "globalize" cleanly.

Methods

The core of the paper is the application of sheaf theory to sensor fusion and tracking. A sheaf is a mathematical structure that organizes data over a topological space, allowing for the representation of relationships and interactions between different data sources.

  • Abstract Simplicial Complex (ASC): Used to model the sensor network, where sensors are vertices and collections of sensors measuring the same quantity form faces.
  • Sheaf of Sets/Vector Spaces: Assigns data (stalks) to each face of the ASC, representing sensor readings.
  • Restriction Maps: Define how data from lower-dimensional faces (individual sensors) combine to form data for higher-dimensional faces (sensor groups).
  • Consistency Structures & Consistency Radius: These concepts are introduced to quantify how well sensor data agrees. The consistency radius provides a global measure of uncertainty, and the consistency filtration breaks down disagreements to identify problematic sensors.
  • Comparison with Dynamic Linear Model (DLM) and Kalman Filter: The authors compare their sheaf-based approach with a traditional statistical method to highlight the strengths of their framework, particularly when data is not pre-processed into a common coordinate system.

Datasets

The study uses experimental data collected from a wildlife tracking experiment involving black bears in Asheville, North Carolina. This dataset includes:

  • GPS readings from bear collars.
  • Radio VHF device data.
  • Text reports from researchers.
  • Vehicle GPS data from the tracking team.
  • "Dummy" bear collars placed at fixed locations.

This dataset is designed to be heterogeneous in terms of data types and error models, making it suitable for testing multi-sensor integration methods.

Key Insights and Ideas

  • Sheaves as a Unified Framework: Sheaf theory provides a mathematically rigorous way to model complex sensor interactions and quantify agreement, offering a more principled approach than ad-hoc fusion methods.
  • Diagnosing Sensor Disagreements: The consistency filtration allows for fine-grained identification of which sensor groups are contributing most to disagreements, offering valuable diagnostic capabilities.
  • Robustness to Unnormalized Data: A key advantage highlighted is the sheaf model's ability to remain valid even when sensor data is not pre-normalized into a common coordinate system, a limitation for traditional DLMs.
  • Theoretical Foundation for Fusion: The paper suggests that any principled sensor integration method will, in some way, recapitulate aspects of sheaf theory, underscoring its fundamental nature.

Key Results

  • The sheaf-based model successfully integrated heterogeneous sensor data to track bears and the human tracking team.
  • The consistency radius analysis revealed a decrease in overall error over time, particularly for bear location estimates, as the human trackers approached the bears.
  • Disagreements were most prominent between the VHF receiver and the bear collar GPS, indicating potential issues with these sensors at certain ranges or conditions.
  • The sheaf model provided a holistic temporal picture of observations and described agreement levels between sensor groups.
  • When data was normalized, the sheaf model's structure mirrored that of the dynamic linear model, demonstrating their underlying structural similarity.

Limitations and Future Work

  • Parameter Estimation: While the sheaf model avoids upfront parameter estimation for sensor self-consistency, the dynamic linear model facilitates this. The authors suggest these approaches are complementary.
  • Dynamic Modeling: The current sheaf model is not inherently dynamic, whereas DLMs explicitly model temporal state changes. The authors note that time can be explicitly topologized within a sheaf framework for future work.
  • Computational Complexity: While not explicitly detailed, the complexity of sheaf computations for very large sensor networks could be a consideration.
  • Interpretability of Consistency Radius: The authors acknowledge that the consistency radius indicates agreement but doesn't inherently tell which measurements are more trustworthy.

Summary updated

Abstract

Integration of multiple, heterogeneous sensors is a challenging problem across a range of applications. Prominent among these are multi-target tracking, where one must combine observations from different sensor types in a meaningful and efficient way to track multiple targets. Because different sensors have differing error models, we seek a theoretically justified quantification of the agreement among ensembles of sensors, both overall for a sensor collection, and also at a fine-grained level specifying pairwise and multi-way interactions among sensors. We demonstrate that the theory of mathematical sheaves provides a unified answer to this need, supporting both quantitative and qualitative data. Furthermore, the theory provides algorithms to globalize data across the network of deployed sensors, and to diagnose issues when the data do not globalize cleanly. We demonstrate and illustrate the utility of sheaf-based tracking models based on experimental data of a wild population of black bears in Asheville, North Carolina. A measurement model involving four sensors deployed among the bears and the team of scientists charged with tracking their location is deployed. This provides a sheaf-based integration model which is small enough to fully interpret, but of sufficient complexity to demonstrate the sheaf’s ability to recover a holistic picture of the locations and behaviors of both individual bears and the bear-human tracking system. A statistical approach was developed in parallel for comparison, a dynamic linear model which was estimated using a Kalman filter. This approach also recovered bear and human locations and sensor accuracies. When the observations are normalized into a common coordinate system, the structure of the dynamic linear observation model recapitulates the structure of the sheaf model, demonstrating the canonicity of the sheaf-based approach. However, when the observations are not so normalized, the sheaf model still remains valid.
BibTeX
@inproceedings{joslyn2020sheaf,
  author = {Cliff A. Joslyn and Lauren E. Charles and Chris DePerno and Nicholas Gould and Kathleen Nowak and Brenda Praggastis and Emilie Purvine and Michael Robinson and Jennifer Strules and Paul Whitney},
  title = {{A Sheaf Theoretical Approach to Uncertainty Quantification of Heterogeneous Geolocation Information}},
  booktitle = {Sensors},
  year = {2020},
  url = {https://openreview.net/forum?id=dWexfAnza4},
}

Loading related content...

Figure