No abstract

No abstract

No abstract

Wall-climbing drones have many applications, including structural health monitoring of civil structures, such as bridges and high-rise buildings, cleaning of solar panels to improve power generation efficiency, and airplane visual inspections. For these applications, the drone requires a high-payload capacity, and consequently the size and weight of the drone increase. The drone also should not damage the target structures considering the purpose of its mission. Our previous versions of a wall-climbing drone could have high-impact force on the surface where the drone perches and on the platform itself because of the impact caused by a fast pose change and landing speed. In order to overcome this potential risk, a mechanism and a control algorithm for perching on a vertical surface through low-speed pose change are proposed in this paper. The drone platform is based on an X-configuration quadcopter, and a tilt-rotor mechanism is incorporated into the two axes, such that the front thrusters and the rear thrusters are paired. The vertical soft landing mechanism using the tilt-rotors is validated by the experimental tests of the prototype.

Among various sensors used to recognize obstacles in marine environments, vision sensors are the most basic. Vision sensors are significantly affected by the surrounding environment and cannot recognize distant objects. However, despite these drawbacks, they can detect objects that radars cannot detect in nearby regions. They can also recognize small obstacles such as boats that are not equipped with an automatic identification system (AIS) or buoys. Thus, vision sensors and radar can be used in a complementary manner. This paper proposes a vision sensor-based model, called Skip-ENet, for recognizing obstacles in real time. Compared with ENet, the amount of computation is not significantly higher. Further, Skip-ENet can segment complex marine obstacles effectively by increasing the values for the class accuracy and mean Intersection of Union (mIoU). Moreover, this model enables even low-cost embedded systems to compute 10 or more frames per second (fps). The superiority of the proposed model was verified by comparing its performance with that of the conventional segmentation models, MobileNet, ENet, and DeeplabV3+.