Abstract
Deep learning-based indoor localization systems often suffer from "sensor perspective shift," where models trained on fixed sensor configurations fail when sensors are relocated. This paper introduces FlexLoc, a novel architecture that achieves zero-shot perspective invariance by employing conditional neural networks. By injecting real-time sensor pose information directly into the model's weights rather than as raw input data, FlexLoc enables systems to adapt to new sensor arrangements without retraining. Experimental results on a multimodal dataset demonstrate a 50% improvement in localization accuracy for unseen sensor perspectives, achieving a mean error of 14.28 cm.
Introduction
High-precision indoor localization is a cornerstone of modern robotics and smart environments, leveraging data from cameras, radar, and acoustic sensors. While deep learning models excel at fusing these multimodal inputs, they are notoriously brittle regarding spatial geometry. Most models implicitly "bake" the specific coordinates and angles of sensors into their neural weights during training. Consequently, moving a sensor for maintenance or reconfiguring a room requires a costly cycle of data collection and retraining.
FlexLoc addresses this bottleneck by decoupling the localization logic from specific sensor placements. The system treats sensor pose (location and orientation) as a dynamic condition that modifies the network's behavior on the fly, allowing for "zero-shot" deployment in entirely new configurations.
Methodology
Conditional Neural Networks for Pose Injection
The core innovation of FlexLoc is the use of Conditional Neural Networks to handle sensor perspective. The authors argue that simply concatenating pose data to the input features is insufficient for capturing complex spatial transformations. Instead, they explore two primary methods for "pose injection":
- Conditional 1D Convolution (CondConv): An auxiliary network processes the sensor's pose (position and quaternion orientation) to generate the weights for 1D convolutional layers. These layers then transform extracted sensor features based on the sensor's specific world-space context.
- Conditional Layer Normalization (CLN): A more computationally efficient approach where the pose data is used to dynamically adjust the scale ($\gamma$) and offset ($\beta$) parameters within the normalization layers.
Multimodal Pipeline
The FlexLoc architecture utilizes specialized backbones—such as ResNet for visual data and Transformers for audio—to extract high-level features. These features are processed by the conditional layers and then fused via a transformer encoder. This fused representation is finally mapped to global coordinates to determine the target's location.
Experimental Setup
The GDTM Dataset
To evaluate FlexLoc, the authors developed the Geospatial Tracking Dataset with Multimodal sensors (GDTM). This dataset comprises nine hours of tracking data across 22 unique sensor arrangements. Each sensor node consists of:
- RGB and Depth cameras
- mmWave radar
- A microphone array
The dataset was split into 13 views for training, 4 for validation, and 5 entirely unseen views to test zero-shot performance. Ground truth was established using a high-precision OptiTrack motion capture system.
Robustness via Sensor Dropout
To ensure the model did not become overly reliant on high-fidelity visual data, the authors implemented a 50% sensor dropout strategy during training. This forced the network to learn localization cues from depth, radar, and audio independently, ensuring robustness in scenarios where specific sensors (like cameras in low light) might fail.
Results and Discussion
Zero-Shot Performance
FlexLoc (specifically the CondConv variant) demonstrated significant superiority over standard early-fusion baselines. In zero-shot scenarios—where the model encountered sensor perspectives it had never seen during training—FlexLoc achieved an average localization error of 14.28 cm. This represents a nearly 50% improvement over the baseline error of 26.86 cm.
Remarkably, FlexLoc’s zero-shot accuracy on unseen perspectives was nearly as high as models trained specifically on those perspectives (~12.9 cm), suggesting that the conditional weights effectively generalize the underlying physics of spatial mapping.
Key Insights
- Weight Modification vs. Input Concatenation: The study confirms that modifying the transformation function (the weights) is a more powerful way to represent geometric relationships than simply adding pose data as an input feature.
- Efficiency: Despite its performance gains, FlexLoc is highly efficient. The CondConv approach adds only 0.2% more parameters, while the CLN approach can actually reduce the total parameter count by replacing standard normalization layers.
Conclusion and Future Work
FlexLoc provides a scalable solution for deploying multimodal localization systems without the need for site-specific retraining. By making models invariant to sensor perspective, it lowers the barrier for real-world robotics applications.
Future research will focus on removing the dependency on external motion capture for sensor poses (e.g., by integrating self-localization via SLAM) and testing the system in more complex, cluttered environments with multiple dynamic targets. Additionally, extending FlexLoc to handle sensors mounted on moving platforms would represent a significant step toward fully autonomous, distributed sensing networks.
Summary updated