Finding Explanations in AI Fusion of Electro-Optical/Passive Radio-Frequency Data

Asad Vakil, Erik Blasch, Robert Ewing, Jia Li

TL;DR

This paper introduces an explainable AI (XAI) framework for fusing electro-optical (EO) and passive radio-frequency (P-RF) sensor data for object detection and tracking. By using a greedy algorithm (explainX.ai) to analyze the weights assigned to different data inputs, the research provides novel insights into how P-RF data contributes to fusion models, enhancing transparency and aiding decision-makers.

What Problem Does This Paper Address?

In many AI applications, especially those using complex neural networks, understanding why a model makes a particular decision is challenging. These "black box" models can be highly effective, but their lack of transparency hinders trust and debugging, particularly in critical domains like defense and autonomous systems. Sensor fusion, combining data from multiple sources like EO cameras and P-RF sensors, offers robustness but can exacerbate this explainability problem. While EO data provides rich visual information, P-RF data, though less intuitive, offers advantages like all-weather operation and stealth. The challenge is to effectively fuse these complementary modalities and, crucially, to understand how the less intuitive P-RF data contributes to the overall decision-making process.

Methods

The core of this research lies in developing an explainable framework for an LSTM-CCA (Long Short-Term Memory - Canonical Correlation Analysis) fusion model.

  • Sensor Fusion Model: An LSTM-CCA model is employed, which leverages Canonical Correlation Analysis (CCA) to find correlations between different data views. This model takes preprocessed P-RF data (histograms of I/Q components) and enhanced EO data (processed with dense optical flow) as inputs for object detection and tracking.
  • Explainability Tool: The explainX.ai tool, based on the ProtoDash algorithm, is used to provide transparency. This greedy algorithm analyzes the weights assigned by the fusion model to different input features (DOF-EO, EO-CCA variates, P-RF histograms, RF-CCA variates) for each scenario and target. This allows researchers to understand the relative importance of each data source in the model's decision-making.

Datasets

The research utilizes the ESCAPE (Experiments, Scenarios, Concept of Operations, and Prototype Engineering) dataset. This dataset is a multimodal benchmark that includes infrared (IR), full motion video (FMV), passive RF, and acoustic data, collected across various scenarios designed for sensor fusion research. The paper specifically uses Scenarios 1, 2(2C), and 3(2D), focusing on the detection and tracking of five types of ground vehicles.

Key Insights and Ideas

  • P-RF Data's Crucial Role: While EO data is often the primary driver for decisions, P-RF data, particularly its CCA variates, plays a significant role, especially when EO data is insufficient (e.g., target out of sight). This highlights the value of P-RF for confirmation and detection.
  • CCA Variates Bridge the Gap: The CCA variates derived from both EO and P-RF data are shown to be highly influential. The RF-CCA variates, in particular, seem to bridge the gap between raw P-RF histograms and accurate target detection, significantly improving fusion performance.
  • Scenario-Dependent Importance: The importance of different data inputs varies significantly depending on the scenario and the specific target being tracked. This emphasizes the need for adaptive fusion strategies.
  • Explainability Enhances Trust: The use of explainX.ai provides concrete weights, demonstrating how the model utilizes each input. This transparency is crucial for understanding the model's behavior and building trust in its predictions.

Key Results

  • The LSTM-CCA fusion model, when analyzed with explainX.ai, achieved a perfect F1 score and tracking detection rate metric of 1.0 for each respective target across the tested scenarios.
  • Analysis of weights revealed that EO Input consistently received the highest weight, but RF-CCA variates were occasionally almost as important as EO inputs, demonstrating their significant contribution.
  • In scenarios where EO data was limited (e.g., Vehicle 5 in Scenario 3(2D)), P-RF CCA input provided enough insight to be weighted equally with EO input.
  • The P-RF histogram input generally received the lowest weight, suggesting that raw P-RF data alone is less informative than its CCA-derived counterpart or EO data.

Limitations and Future Work

  • Limited Scenarios and Targets: The research focused on a subset of the ESCAPE dataset (Scenarios 1, 2(2C), and 3(2D)) and a limited number of vehicle targets. The findings might not generalize directly to more complex environments or a wider variety of targets.
  • Data Frame Fusion Model Expansion: The current explainability framework is applied to a specific data frame fusion model. Future work could involve expanding this framework to more complex models with a greater number of input features.
  • Deeper Dive into Local Interaction: While the paper focuses on global explanations, the authors suggest that exploring local interaction aspects of explainability could provide further insights, especially as the model complexity increases.
  • Investigating Other XAI Techniques: The paper mentions other XAI techniques like activation maximization, which could be explored for deeper understanding, particularly for scenarios where P-RF data shows unexpected behavior.

Summary updated

Abstract

In the Information Age, the widespread usage of blackbox algorithms makes it difficult to understand how data is used. The practice of sensor fusion to achieve results is widespread, as there are many tools to further improve the robustness and performance of a model. In this study, we demonstrate the utilization of a Long Short-Term Memory (LSTM-CCA) model for the fusion of Passive RF (P-RF) and Electro-Optical (EO) data in order to gain insights into how P-RF data are utilized. The P-RF data are constructed from the in-phase and quadrature component (I/Q) data processed via histograms, and are combined with enhanced EO data via dense optical flow (DOF). The preprocessed data are then used as training data with an LSTM-CCA model in order to achieve object detection and tracking. In order to determine the impact of the different data inputs, a greedy algorithm (explainX.ai) is implemented to determine the weight and impact of the canonical variates provided to the fusion model on a scenario-by-scenario basis. This research introduces an explainable LSTM-CCA framework for P-RF and EO sensor fusion, providing novel insights into the sensor fusion process that can assist in the detection and differentiation of targets and help decision-makers to determine the weights for each input.
BibTeX
@inproceedings{vakil2023finding,
  author = {Asad Vakil and Erik Blasch and Robert Ewing and Jia Li},
  title = {{Finding Explanations in AI Fusion of Electro-Optical/Passive Radio-Frequency Data}},
  booktitle = {Sensors},
  year = {2023},
  url = {https://openreview.net/forum?id=Klf2NXBN0fE},
}

Loading related content...

Figure