A Comparative Study on Segmentation Techniques for Context-Aware Safe Landing of UAVs

As the use of Unmanned Aerial Vehicles (UAVs) in various tasks within human-inhabited environments becomes increasingly common, critical aspects such as emergency landing need to be addressed. The use of deep learning has become widely adopted to provide context-sensitive solutions, where semantic segmentation has shown promising results. Therefore, this paper presents a comparative study that aims at evaluating several candidate semantic segmentation algorithms, offering a guide for an appropriate selection of the segmentation module in a UAV safe landing task in unstructured urban environments. More specifically, a comparison was made between three prominent segmentation models, U-Net, SegFormer, and MANet, using the Semantic Drone Dataset. First, the models were evaluated using the original 24 classes of the dataset. SegFormer performed slightly better than the other algorithms tested. In a second, more critical experiment, the classes were grouped into six risk levels based on the Specific Operations Risk Assessment (SORA) framework. To address class imbalance and prioritize high-risk categories, a weighted Cross-Entropy loss was employed, assigning higher penalties to misclassifications in critical risk levels. In this setup, MANet achieved the best results, showing its ability to adapt to risk-based classification and capture important features. Later, the three models were optimized to prove that even complex architectures can be enhanced for real-time inference. The optimization included converting the model to TensorRT format and using FP16 precision, which reduced the models’ size by at least 30%. Finally, the optimized U-Net, the largest model, was tested on a NVIDIA Jetson Orin Nano platform achieving real-time inference. This shows that heavy models can run on embedded hardware like the Jetson Orin Nano, making them viable for safe, autonomous UAV applications.