An Approach to Visual Odometry Based on End-to-End Semi-Supervised Generative Adversarial Networks

Abstract

A method of visual odometry based on an end-to-end semi-supervised generative adversarial network, involving a visual odometry algorithm. Build a feature generation network; build a discriminant network; confront training; solve dynamic scene problems. The SIFT feature algorithm is used to annotate point position information and extract feature descriptors, and then use the randomly generated homography matrix to obtain relevant frames and match feature points to generate corresponding training labels. The generation network can input the original image to generate the corresponding feature point position and corresponding depth description. The discriminative network combines the semantic-geometric consistency loss function, the feature point cross-entropy loss function and the discriminative loss function to compete with the generative network. Through the training of GAN, the generative network can generate point position information and depth description that make the discriminative network indistinguishable, thus avoiding manual design of local features.

Description

Technical field [0001] The present invention relates to visual odometry calculation method, particularly to a method based on the generated end against the semi-supervised network the visual odometry. Background technique [0002] In the past few decades, mobile robots and autonomous driving technology field attracted wide attention from researchers worldwide, we have made significant progress and breakthroughs. At present, the mobile robot can autonomously perform complex tasks, such as research and development engineering company Boston Dynamics robot has been able to imitate human make backflips, open the door and other activities. Autopilot technology has also made a major breakthrough, automatic driving car is expected to mass production within two years. Whether mobile robots or automated driving technology needs in a complex and dynamic environment, indoor or outdoor automatic navigation. To enable autonomous navigation, navigation carriers need to locate itself in the sur...

AI Technical Summary

Problems solved by technology

At present, this supervised learning method has the following difficulties: First, it is extremely difficult to obtain data with a large amount of geometric information annotations. One hundred thousand images are tens of millions or even hundreds of millions of points, so it is very difficult to label such a labor-intensive

Therefore, some researchers, such as Kendall et al. [2] Using Visual SFM with Transfer Learning [3] Marking the image pose saves a lot of time and labor costs, but the accuracy is also limited by the Visual SFM algorithm

Secondly, if you only learn the pose of a single image corresponding to the current reference coordinate system, then the usage scenarios will be greatly limited

Images