ENGLISH / MAGYAR
Kövess
minket

Deep Learning-Based Multi-Modal Narrow Road Extraction for UAV and Satellite Imagery

2024-2025/II.
Liu Chang

MSc Thesis Topic: Deep Learning-Based Multi-Modal Narrow Road Extraction for UAV and Satellite Imagery

Description:

Extracting narrow roads from high-resolution aerial and satellite imagery is a critical challenge in applications such as autonomous navigation, urban planning, and disaster management. Narrow roads often appear occluded, discontinuous, or embedded in complex environments, making their detection significantly more difficult than broader road networks. While deep learning has shown promise in semantic segmentation tasks, traditional models frequently struggle with narrow and intricate road structures due to their limited ability to capture fine-grained details.

This research focuses on leveraging multi-modal data fusion from UAV (Unmanned Aerial Vehicle) and satellite imagery to improve narrow road detection. UAV imagery provides high-resolution, low-altitude perspectives, while satellite imagery offers broader spatial context. By combining these complementary modalities, this thesis proposes a novel deep learning-based model designed to effectively detect narrow roads in diverse and challenging environments.

Objectives:

Model Development for Narrow Road Detection: Develop a novel deep learning architecture optimized for detecting narrow roads in complex environments. Incorporate advanced techniques such as dilated convolutional layers and attention mechanisms to improve feature extraction and segmentation accuracy for narrow and occluded roads.


Multi-Modal Data Fusion: Fuse data from UAV and satellite imagery to integrate high-resolution detail and large-scale context. Design a pipeline for pre-processing and aligning multi-modal datasets to enable efficient model training and inference.


Performance Evaluation: Conduct rigorous evaluations of the proposed model using standard metrics, including Intersection over Union (IoU), F1-score, and precision-recall curves. Compare the model's performance against existing baseline methods to demonstrate its advantages, particularly in narrow road detection. Analyze the impact of multi-modal data fusion on detection accuracy and robustness in various environments.

The applicant will use publicly available datasets and receive technical support from SZTAKI.


Expected Contributions:

This thesis aims to advance the state of narrow road detection by:

Proposing a novel deep learning model tailored to the challenges of narrow and occluded roads.
Demonstrating the effectiveness of multi-modal data fusion for integrating high-resolution UAV imagery and broad-context satellite imagery.
Providing insights into how advanced deep learning techniques can improve road extraction accuracy, especially in challenging scenarios.
The findings of this research are expected to benefit real-world applications such as autonomous UAV navigation, disaster response, and transportation system optimization.

Requirements and Technical Skills:

Proficiency in Python and experience with deep learning frameworks such as TensorFlow or PyTorch.
Familiarity with semantic segmentation models and techniques for data fusion.
Knowledge of image pre-processing, augmentation, and machine learning evaluation metrics.
Ability to work independently on experimental design and implementation.


1
1