Design method of monocular visual odometer based on unsupervised deep learning

• Main Authors: CHEN YANGZHUO, LI PENG, ZHANG YING, PAN HONGBIN, DOU JIE, MENG BUMIN, CAI CHENGLIN, ZHOU YAN, HUANG PENG, LI XIMIN, CAI XIAOWEN
• Format: Patent
• Language: Chinese; English
• Published: 01.07.2022
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; GYROSCOPIC INSTRUMENTS; IMAGE DATA PROCESSING OR GENERATION, IN GENERAL; MEASURING; MEASURING DISTANCES, LEVELS OR BEARINGS; NAVIGATION; PHOTOGRAMMETRY OR VIDEOGRAMMETRY; PHYSICS; SURVEYING; TESTING

The invention discloses a monocular visual odometer design method based on unsupervised deep learning, which improves the performance of a visual odometer by jointly training depth, relative attitude and optical flow, and obtains depth information and dense optical flow information with consistent long sequences by using a deep network and an optical flow network. Accurate sparse optical flow sampling is carried out through a front-back consistency error, an optimal tracking mode is selected according to model scores and is aligned with depth information to obtain a visual odometer with consistent scale, and a visual odometer with consistent scale is obtained by combining geometric constraint conditions of a traditional method and robustness matching of a deep network. The method is obviously superior to a pure geometric method and an end-to-end deep learning method in multiple error evaluation indexes, and experiments prove that the method effectively reduces the problems of scale inconsistency and scale dri