3D scene geometry estimation method of substation inspection robot based on lightweight neural network

Abstract

Understanding 3D scene geometry from video is a basic subject of visual perception. It includes many classic computer vision tasks, such as depth recovery, traffic estimation, visual odometer. Recent work has proved that deep learning can be applied to scene understanding problems. But they all have some inherent limitations. For example, they need stereo cameras as additional devices for data acquisition, or can't explicitly deal with non-rigid and occlusion. The environment in the substation is complex, and there are many devices. In the working process of inspection robot, the target is very easy to be blocked, and it is difficult to deploy directly by traditional methods. In addition, the real-time performance of neural network is very important for electric inspection robot. In this paper, 3D scene geometry estimation method of substation inspection robot is proposed, which consists of two main parts: GeoNet module and pruning module. Experiments show that the proposed method can be effectively applied to electric inspection robot.