Previous Abstract Return to Session C6

Session C6: Terrestrial Signals of Opportunity-Based Navigation Systems

Radio Propagation Learning: Predicting Reflection Points from Radio Attenuation Maps
Martin Schmidhammer, Benjamin Siebler (German Aerospace Center (DRL), Institute of Communications and Navigation, Ulrich Hillenbrand, DLR, Institute of Robotics and Mechatronics

The accurate prediction of reflection points (RPs) in radio propagation is critical not only for optimizing wireless communication systems but also for enhancing radio-based sensing applications. In multipath environments, signals reflecting off various surfaces introduce distinct propagation paths that can provide useful spatial and structural information about the environment. Identifying these reflection points is thus essential for applications in radio tomographic imaging (RTI) and device-free localization (DFL), which utilize signal fading to sense changes and positions within a space without requiring active devices on tracked entities.
This paper presents a novel, measurement-driven approach for predicting the locations of RPs leveraging artificial intelligence (AI). Unlike traditional methods that require detailed knowledge of the environment, such as floor plans, our AI-driven model is capable of directly inferring RP locations from power changes induced by users. Additionally, we introduce a model-based approach for comparison in simpler scenarios with single-bounce reflections, enabling us to assess the feasibility of each method in practical applications.
Our methodology consists of two main approaches for predicting RPs: an AI-based neural network model and a model-based optimization approach, each leveraging distinct attributes of multipath signals to infer locations of RPs.
1. AI-Based Approach: The AI approach employs a deep neural network to predict RP locations based on the encoded scene data. Key components include:
a. Scene Encoding: A grid of 33x33 basis points is superimposed on the area where RPs could exist, with each point’s distance to the nearest data point and corresponding attenuation values forming a 2178-dimensional vector encoding for each scene. This straightforward encoding structure provides a rich representation of the environment while leaving the learning of spatial relationships to the higher network layers.
b. Neural Network Architecture and Ensemble Prediction: For RP prediction, we implemented a deep fully connected neural network with progressively smaller layers to regress the coordinates of one, two, or more RPs. The deep network architecture scales down to the output layer, as needed for the respective order of reflections. To increase prediction accuracy, an ensemble of neural networks was trained with different initializations. Statistics of the ensemble results achieves robust predictions and gives intuition about the expected prediction accuracy. Local optimization based on known constraints (e.g., distance and path length) refines the final predicted RP locations.
c. Training and Optimization: Each network in the ensemble was trained using the ADAM optimizer with fine-tuned parameters for optimal convergence. Training concluded upon reaching stable validation performance, ensuring high prediction accuracy across the ensemble.
2. Model-Based Approach: For comparison, a model-based approach targets first-order reflections, solving an optimization problem based on an empirical fading model and Gaussian noise assumptions. This method constrains possible RP locations to the elliptic arc according to the propagation delay, providing an efficient solution for simpler scenarios.
Dataset Generation and Scene Setup: We use simulated datasets to evaluate and train the AI model under realistic multipath propagation conditions. These datasets represent random transmitter and receiver configurations in a 48m by 48m area with randomly distributed RPs and user trajectories simulated along 2D waypoints. The scenes include signal attenuation data at points close to direct signal paths, transmitter and receiver positions, and ground-truth locations of the RPs for training and validation purposes. A coordinate transformation aligns the transmitter and receiver symmetrically, reducing scene variability and facilitating efficient model learning.
The AI-based approach was anticipated to offer effective RP prediction for both first- and higher-order reflections, a hypothesis confirmed through simulation results. Meanwhile, the model-based method was expected to work well with single reflections but fall short with higher-order cases due to its simplified assumptions.
- AI-Based Approach: The deep neural network ensemble achieved highly accurate RP location estimates across diverse configurations and trajectories. It successfully handled higher-order reflections, where traditional model-based methods cannot operate, and provided a quantifiable measure of uncertainty through ensemble dispersion.
- Model-Based Approach: The model-based method produced reliable predictions for first-order reflections under constrained conditions, serving as a viable solution for single-bounce-reflection scenarios with Gaussian noise and limited multipath complexity.
This study demonstrates the viability of using AI for RP localization, delivering accurate predictions without relying on prior environmental layouts. In addition to improving wireless communication applications, our results highlight the potential for this approach in radio sensing tasks, such as RTI and DFL. Key takeaways include:
- Applicability to Fading-Based Radio Sensing: By estimating RPs, our approach provides valuable information for radio sensing methods that rely on signal fading, such as RTI and DFL. These applications can leverage RP locations to enhance spatial resolution, detect movement, and interpret environmental structure in real-time, making AI-based RP prediction a significant asset in non-invasive sensing.
- Scalability for Complex Multipath Environments: The AI approach is scalable for environments with multiple reflections, offering robust prediction where model-based methods are infeasible. This capability allows broader use in dynamic environments common in IoT, smart homes, and industrial sensing.
- Ensemble Prediction and Uncertainty Quantification: The ensemble approach not only refines prediction accuracy but also provides a measure of prediction reliability, which is crucial for applications where high-confidence positioning and sensing are required.
The proposed method advances both wireless communication and radio-based sensing by introducing a scalable, environment-agnostic model for RP localization. By enabling accurate RP estimation without floor plans or extensive pre-mapping, our approach offers substantial benefits for fading-based sensing techniques like RTI and DFL, enhancing their ability to detect and localize objects or individuals within a space. In wireless communications, our model improves indoor positioning, network planning, and interference management. In radio sensing, the RP predictions facilitate refined object localization and environmental understanding, promising advancements in applications ranging from autonomous navigation and smart infrastructure to health monitoring and security. This work thus positions AI-based RP prediction as a core technique in the evolving landscape of multipath-aware sensing and communication systems.

Previous Abstract Return to Session C6