Vision-Based Autonomous UAV Navigation and Landing for Urban Search and Rescue

TL;DR

This work tackles autonomous UAV navigation and landing for urban search-and-rescue in unknown post-disaster environments where pre-existing maps are unreliable. It introduces a pixel-level landing-site detection pipeline that fuses hazard costmaps for depth confidence ($J_{DE}$), flatness ($J_{FL}$), steepness ($J_N$), and energy ($J_{EC}$) into a final score $J = c_1 J_{DE} + c_2 J_{FL} + c_3 J_N + c_4 J_{EC}$ with $c_i \in [0,1]$ and $\sum c_i = 1$, followed by global clustering to produce sparse, safe landing regions; a minimum-jerk trajectory ensures smooth landings. The system integrates state estimation (ROVIO+EKF), online mapping with Octomap and Voxblox TSDF, and a next-best-view planner to explore unknown debris-filled environments, all running on a low-power onboard computer. The authors also introduce AutoLand, a first-of-its-kind synthetic dataset with over $1.2$ million RGB images plus groundtruth depth, normals, semantics, and camera poses, and validate the approach in both hyperrealistic simulation and real-world collapsed buildings, showing robust landing-site detection and practical, time-critical performance for USAR operations.

Abstract

Unmanned Aerial Vehicles (UAVs) equipped with bioradars are a life-saving technology that can enable identification of survivors under collapsed buildings in the aftermath of natural disasters such as earthquakes or gas explosions. However, these UAVs have to be able to autonomously navigate in disaster struck environments and land on debris piles in order to accurately locate the survivors. This problem is extremely challenging as pre-existing maps cannot be leveraged for navigation due to structural changes that may have occurred. Furthermore, existing landing site detection algorithms are not suitable to identify safe landing regions on debris piles. In this work, we present a computationally efficient system for autonomous UAV navigation and landing that does not require any prior knowledge about the environment. We propose a novel landing site detection algorithm that computes costmaps based on several hazard factors including terrain flatness, steepness, depth accuracy, and energy consumption information. We also introduce a first-of-a-kind synthetic dataset of over 1.2 million images of collapsed buildings with groundtruth depth, surface normals, semantics and camera pose information. We demonstrate the efficacy of our system using experiments from a city scale hyperrealistic simulation environment and in real-world scenarios with collapsed buildings.

Figures (9)

• Figure 1: Left: Our UAV autonomously navigating in an environment with collapsed buildings attempting to identify trapped survivors. Middle: A 3D volumetric reconstruction of the scene, additionally illustrating the minimum-jerk trajectory to the detected safe landing site. Right-Top: The corresponding full costmap computed by our landing site detection algorithm. Right-Bottom: Corresponding dense candidate landing sites that were detected projected onto the local 3D map.
• Figure 2: Overview of our autonomous navigation and landing system. We use the DJI M100 quadrotor with the N1 flight controller, the NVIDIA Jetson TX2 for computation, and a ZED stereo camera for acquiring depth information. All our mapping, localization and landing site detection algorithms run online on the Jetson TX2.
• Figure 3: Overview of our landing site detection algorithm. The figure illustrates the various costmaps for hazard estimation. Scale: Red indicates highest assigned score while blue indicates the lowest score. The detected landing sites are projected on to a volumetric 3D reconstructed mesh of the environment.
• Figure 4: Example images from our hyperrealistic synthetic dataset showing the RGB image with the corresponding groundtruth depth, surface normals and pixel-level semantic labels for nine object categories. Note that the depth image and the surface normals are colorized only for visualization.
• Figure 5: (a): Comparison of the path generated using a strictly sampling-based RRT* approach with a polynomial steer function (blue line) and joint polynomial optimization method (red line). (b): Trajectory generated by the lawn-mower scan that was traversed by our UAV. (c): Trajectory generated by the next-best-view planner that traversed by our UAV.