In non-cooperative environments, unmanned aerial vehicles (UAVs) have to land
without artificial markers, which is a key step towards achieving full autonomy.
However, the existing vision-based schemes have the common problems of poor
robustness and generalization, and the LiDAR-based schemes have the
disadvantages of low resolution, high power consumption and high weight. In this
paper, we propose an UAV landing system equipped with a binocular camera to
preform 3D reconstruction and select the safe landing zone. The whole system
only consists of a stereo camera, and the innovation of the solution is fusing
the stereo matching algorithm and monocular depth estimation(MDE) model to get a
robust prediction on the metric depth. The whole landing system consists of a
stereo matching module, a monocular depth estimation (MDE) module, a depth
fusion module, and a safe landing zone selection module. The stereo matching
module uses Semi-Global Matching (SGM) algorithm to calculate the binocular
disparities to get the dense metric depth of each pixel, and is deployed on GPU
to meet the real-time requirements. The MDE module conducts relative depth
estima-tion on the left-eye image, and is also deployed on GPU to improve the
inference speed. The output of the MDE module has the advantages of high
accuracy and excellent generalization but it is ambiguous in scale, and that of
the stereo matching module is deterministic in scale but susceptible to
illumination and moving objects. Considering the above features, the depth
fusion module fuses the relative depth estimation result and the metric depth
information into a robust and accurate metric depth map. With the metric depth
map and camera intrinsic parameters, the safe landing zone selection module
calculates the first and second order derivatives to detect the obstacles and
finds the safe landing zone. In AirSim, we build a UAV Hardware-In-the-Loop
(HIL) simulation system, and carry out a series of autonomous landing
experiments. The results show that our landing scheme performs 3D reconstruction
of the landing terrain and selects safe landing zone with high efficiency and
reliability.