Vision-Based Risk Aware Emergency Landing for UAVs in Complex Urban Environments

TL;DR

The paper tackles emergency UAV landing in complex urban environments by introducing a vision-based, risk-aware framework that relies on a downward monocular camera. It converts semantic segmentation into risk maps, maintains a persistent risk memory, and adaptively expands risk regions by altitude to identify a Safe Landing Zone (SLZ). A local-minimum search combined with temporal stabilization selects robust landing points, while a PD-based control scheme guides the UAV to descend safely. Experimental validation on real urban footage demonstrates high success rates under controlled conditions and highlights the method's potential to enable reliable, markerless emergency landings in unstructured cityscapes.

Abstract

Landing safely in crowded urban environments remains an essential yet challenging endeavor for Unmanned Aerial Vehicles (UAVs), especially in emergency situations. In this work, we propose a risk-aware approach that harnesses semantic segmentation to continuously evaluate potential hazards in the drone's field of view. By using a specialized deep neural network to assign pixel-level risk values and applying an algorithm based on risk maps, our method adaptively identifies a stable Safe Landing Zone (SLZ) despite moving critical obstacles such as vehicles, people, etc., and other visual challenges like shifting illumination. A control system then guides the UAV toward this low-risk region, employing altitude-dependent safety thresholds and temporal landing point stabilization to ensure robust descent trajectories. Experimental validation in diverse urban environments demonstrates the effectiveness of our approach, achieving over 90% landing success rates in very challenging real scenarios, showing significant improvements in various risk metrics. Our findings suggest that risk-oriented vision methods can effectively help reduce the risk of accidents in emergency landing situations, particularly in complex, unstructured, urban scenarios, densely populated with moving risky obstacles, while potentiating the true capabilities of UAVs in complex urban operations.

Figures (5)

• Figure 1: Left: Drone-view image $I_k$ from the UAV. Right: Grayscale representation of the risk map $\mathcal{R}_{k}$, where brighter pixels denote higher risk.
• Figure 2: Two consecutive image frames demonstrating the risk memory approach. Column (a) shows aerial drone views $I_{k-1}$ and $I_k$ with moving vehicles. Column (b) displays the raw risk maps $\mathcal{R}_{k-1}$ and $\mathcal{R}_{k}$ generated from semantic segmentation. Column (c) illustrates the corresponding local views of the accumulated global risk maps $\mathcal{R}^{\text{global}}_{k-1}$ and $\mathcal{R}^{\text{global}}_{k}$ as seen from the drone's current perspective, showing how the pixel-wise maximum operation effectively preserves high-risk regions from previous frames within the drone's field of view.
• Figure 3: Progressive visualization of the risk expansion process. Top row shows 2D representations and bottom row shows corresponding 3D intensity profiles of the following. (a) Local view of the risk map $\mathcal{R}^{\text{local}}_{k}$. (b) Map after morphological dilation $D_{k_d}(\mathcal{R}^{\text{local}}_{k})$. (c) Final processed risk map after Gaussian filtering $\mathcal{R}^f_{k}$.
• Figure 4: Visualization of the landing point selection process. (a) Filtered risk map $\mathcal{R}^{\text{f}}_k$ with darker areas representing lower risk. (b) Distance map $\mathcal{L}$ centered at the image's optical center. (c) Weighted combination map $\mathcal{V}$ incorporating both risk and distance factors, with the white dot indicating the selected landing point $p^*_k$ that minimizes this combined cost. (d) Original drone view $I_k$ with the landing point projected onto the actual scene, where a road was selected, which, being free of traffic at that moment, constituted a valid and safe landing point.
• Figure 5: Visualization of the landing point stabilization process at different altitudes. The blue dot represents the instantaneous proposed landing spot $p^*_k$, the green dot shows the temporal averaging of landing points $\overline{p^*_k}$, the red dot marks the center $\mathbf{c}$ of image $I_k$, and the yellow circle is the the safety area $\mathcal{S}$. (a) Initial approach at higher altitude. (b) Descent phase showing the UAV's path toward the selected landing point. (c) and (d) lower altitude views.