Compact Convolutional Segmentation for Visual Landmark Extraction in GNSS-Denied UAV Navigation

Technical Summary: Compact Convolutional Segmentation for Visual Landmark Extraction in GNSS-Denied UAV Navigation

Problem Statement
The reliable localization of Unmanned Aerial Vehicles (UAVs) is severely compromised when Global Navigation Satellite System (GNSS) signals are degraded, blocked, or jammed. While inertial navigation and LiDAR offer alternatives, they often suffer from drift accumulation, high hardware costs, or system complexity. Visual navigation using onboard cameras presents a practical alternative, specifically through the identification of spatially stable and distinctive landmarks. However, extracting these landmarks from aerial imagery is challenging due to scale variations, shadows, background clutter, and viewpoint-dependent changes. The core problem addressed is the lack of a dedicated, lightweight front-end architecture capable of segmenting candidate visual landmarks (specifically stable man-made structures like buildings) from aerial imagery to support subsequent matching or localization stages in GNSS-denied environments.

Methodology
The study proposes a compact, fully convolutional segmentation framework designed as a visual front-end rather than a complete navigation system. The methodology involves the following components:

• Architecture Design: The model utilizes a U-shaped encoder-decoder structure. It integrates three key mechanisms:
  1. Dilation-based Spatial Context: Dilated convolutions are employed to enlarge the effective receptive field without significantly increasing parameters, allowing the capture of multi-scale spatial structures.
  2. Residual Feature Transfer: Residual connections are incorporated to preserve earlier feature information and mitigate degradation in deeper networks, supporting spatial detail recovery after downsampling.
  3. Compact Channel Configuration: Unlike standard architectures using powers of two, the model employs a non-standard, compact channel configuration inspired by Fibonacci and prime numbers (5, 23, 89, and 233 channels) to maintain a lightweight footprint suitable for potential onboard constraints.
• Dataset and Training: Due to the unavailability of a dedicated UAV landmark dataset, the authors adapted the Mnih aerial building segmentation dataset (Boston and Massachusetts regions). Original 1500×1500 images were partitioned into 500×500 sub-images and resized to 192×192 pixels. The dataset comprised 600 training, 63 validation, and 26 test images.
• Training Configuration: The network was trained using the Adam optimizer (learning rate 0.005) with a batch size of 200 for 10 epochs. Batch normalization and dropout were applied for regularization.

Key Contributions and Experimental Evaluation
The primary contribution is the proposal and evaluation of a specific compact architecture combining dilation and residual learning for landmark extraction. The study conducts an ablation-style comparison of four model variants under identical training conditions:

• Model 0 (Plain): A baseline convolutional network.
• Model 1 (Dilation): Adds dilated convolutions.
• Model 2 (Residual): Adds residual connections.
• Proposed Method: Combines both dilation and residual mechanisms.

Results
Experimental results indicate that the proposed architecture achieves the lowest loss on both training and test sets.

• Training Performance: The proposed method yielded the highest Intersection over Union (IoU), precision, and recall among all variants on the training set.
• Test Performance: On the test set, Model 2 (Residual only) achieved higher IoU, precision, and recall values than the proposed method, though the proposed method maintained the lowest test loss.
• Qualitative Analysis: Visual predictions show the model successfully responds to building-like structural patterns while suppressing non-landmark regions (water, roads). However, the results also highlight limitations, including relatively low recall values, boundary sharpness issues, and incomplete region detection.

Significance and Claims
The paper modestly positions its findings as a feasibility study for a visual front-end rather than a finalized navigation system.

• Feasibility: The study demonstrates that segmentation-based feature learning can serve as a viable initial step for extracting candidate visual landmarks from aerial imagery.
• Architectural Insight: The combination of dilation and residual processing provides a stable, compact architecture, though the specific benefits of residual connections appear more pronounced for spatial information preservation in this specific configuration.
• Limitations and Future Work: The authors explicitly state that the current results are not a state-of-the-art benchmark. The low recall and generalization issues are attributed to the limited number of training epochs, the small test set, and the lack of UAV-specific data. The paper concludes that further work is required, including extended training, the creation of UAV-specific annotated datasets, and integration with matching or localization algorithms. Additionally, while the design is lightweight, the authors note that real-time onboard deployment suitability requires further validation through inference-time, memory, and energy measurements on embedded hardware.