Disparity estimation method of electric inspection robot based on lightweight neural network

Abstract

The image depth information can be used to understand the geometric relationship of image scenes, and has important applications in robots, scene understanding, three-xdimensional reconstruction and other fields. Recent work has proved that depth estimation from stereo RGB image pairs can be realized by convolution neural network. However, the mainstream depth estimation deep learning algorithms rely on patch-based Siamese networks, which lacks the ability to comprehensively utilize the context information and the environment texture information, and their performance is poor in complex regions and illposed regions. The working environment of electric inspection robot is complex and the occlusion problem is serious, so the classical methods are difficult to be directly applied. In addition, the real-time performance of neural network is very important for electric inspection robot. In this paper, a disparity estimation neural network for electric inspection robot is proposed, which consists of two main parts: PSMNet module and lightweight cutting module. Experiments show that the proposed method can be effectively applied to electric inspection robot.