Unsupervised Domain Adaptation for Environmental Recognition in Crane Operations

In loading work with a helicopter or a crane, the body itself recognizes an environment, and in order to develop a system that judges whether the commands and operations by human are safe, it needs to realize semantic recognition of the environment, distance recognition to an obstacle, and classification and pursuit of moving objects with high precision using sensors. A method of realizing semantic segmentation, depth estimation and optical flow simultaneously from a camera image had been proposed with a multitasking DNN that took account of the posture and speed of a drone considering a loading work in the air, and it was proved to be useful in a simulated environment. Note however that the model learned in the simulation environment is not a thing suitable for the environmental recognition in a real world. Therefore, this paper aims to develop an environmental recognition system that can be used in the real world by conducting a domain adaptation with adversarial learning. The usefulness of the domain adaptation technique in the proposed multitasking DNN is verified by carrying out environmental recognition of the actual image acquired from the boom tip of a crane.