Terrain-based landing site selection and path planning for fixed-wing UAVs

Computer Vision for Autonomous UAV Flight Safety: An Overview and a Vision-based Safe Landing Pipeline Example

“…Landing site detection methods exploiting a-priori terrain information are actually the most prominent. The authors in [42] detect landing sites for fixed-wing UAVs in emergency situations by using the average height and height variance inside quadtree-based DEM partitions. Partitions whose height variance is below a limit are selected as landing sites and merged with neighboring ones, if they have similar average heights.…”

“…Pros Cons [50] three-category safety labels, relies only on RGB camera input validated on a custom dataset instead of a well-known public one, relies only on obsolete algorithms/methods [27] 3D reconstruction of the flight environment, relies only on a moving RGB monocular camera instead of requiring a 3D rig, 3D point cloud modeling Structure-from-Motion(SfM) and disparity map creation prone to approximation environment modeling errors [49] color and texture features are taken into account for determining landing sites, relies only on RGB camera input only two-category safety labels, relies only on obsolete algorithms/methods [55] multimodal functional state checking from a variety of sensors, 3D reconstruction of the flight environment 3D reconstruction prone to approximation environment modeling errors, relies only on obsolete algorithms/methods [91] relies only on RGB camera input only detects unobstructed areas (one-category safety label), relies only on obsolete algorithms/methods [86] lightweight CNN-based, 3D reconstruction of flight environment, relies only on RGB camera input, three-category plane orientation/geometric labels does not extract scene semantics, two-category final safety labels, [42], [14], [61], [121] exploits a-priory terrain information in the form of Digital Elevation Models (DEMs), simple geometric analysis of terrain roughness and slope relies only on obsolete algorithms/methods [88] 3D CNN-based, takes into account multimodal properties (constraints and atmospheric conditions) increased computational and system complexity, requires LiDAR sensors input of the trained system is a stream of globally registered point clouds in conjunction with a predefined candidate region of interest. The output is a probabilistic safety prediction indicating the safe landing zones.…”

“…Exploiting the 3D Map and the UAV/Camera 3D Pose. Firstly, [42] uses two path planners (Rapidly Exploring Random Tree, or RRT [33], and Particle Swarm Optimization, or PSO [116]) for path planning in limited time from the current UAV position to the closest detected landing site. RRT is a method that systematically connects a goal point with a starting reference point while avoiding rigid obstacles, by expanding the nodes of a tree trajectory generating random points which are uniformly distributed all over the free space.…”

“…[64] supports multiple-robot fleets, high accuracy requires external self-localization system (e.g., GPS+IMU) and known landing spot coordinates, dynamic memory allocation may easily lead to memory overflows, supports only 2D map [116], [67], [42] high accuracy requires external localization system (e.g., GPS+IMU) and known landing spot coordinates, stochasticity in the obtained solution [27] exploits multiple information sources (distance, angular velocity, jerk) for increased accuracy, 3D world coordinates of the landing spot do not need to be known, relies only on RGB camera input to visually detect landing spot, does not depend on external self-localization systems (suitable for GPS-denied environments) increased complexity due to minimization of multiple cost functions, prone to errors [7] does not depend on external self-localization systems (suitable for GPS-denied environments) relies only on RGB camera input to visually detect landing spot, supports moving landing spots (UGV-mounted) increased complexity of control laws, prone to errors [124], [34] relies only on RGB camera input to visually detect landing spot, supports moving landing spots (UGV-mounted) requires external self-localization system (e.g., GPS+IMU), increased complexity of control laws A SLAM algorithm is utilized in order to on-the-fly map the area near the landpad, so that a proper trajectory for safe landing can be designed and executed.…”

Safe landing site detection using SRTM data for the Unmanned Aerial Vehicles

“…yararlanılması, Mayank Garg ve arkadaşları[4], yaptıkları çalışmada İnsansız Hava Araçları için arazi bilgilerine dayalı, iniş alanı seçimi ve güzergâh planlama sistemi geliştirmiştir. İHA'ların görev uçuşlarını sürdürürken, elektronik komponentlerinin arızalanması veya dış ortamlardan kaynaklanan arızalanmalarında iniş için gerekli algoritma geliştirilmiştir.…”