Training of visual positioning network, control method and device of unmanned vehicle

Abstract

This specification discloses a visual positioning network training and a control method and device for unmanned driving equipment. When the feature-regression network is trained, it is not necessary to label the actual road images, and it can be directly based on the output of the feature-regression network. The output results are self-supervised training, which effectively improves the efficiency of feature-regression network training. Moreover, since the network weights of the trained feature-regression network can be extracted, and the weights of the feature extraction branch network included in the visual localization network can be initialized, this can not only effectively improve the target recognition ability of the visual localization network, but also through The way of initial network weight can greatly improve the training efficiency of visual localization network.

Description

technical field [0001] This specification relates to the field of unmanned driving technology, in particular to a method and device for training a visual positioning network and controlling unmanned driving equipment. Background technique [0002] At present, visual positioning technology is widely used in various fields. In the technical field of unmanned driving, after the unmanned driving equipment adopts the image, the position based on which the unmanned driving equipment collects the image can be determined through the visual positioning technology. Attitude data, and based on the pose data, make relevant decisions such as path planning and unmanned equipment control. [0003] In the prior art, the visual localization method based on deep learning is usually implemented by using a trained visual localization network. That is, the collected images are input into the trained visual positioning network to obtain the pose data based on which the unmanned driving device co...

AI Technical Summary

Problems solved by technology

[0004] However, in the above training process, a large number of semantically marked training samples are usually required to train the visual localization network, and the labeling work will inevitably consume a lot of time and cost, thereby increasing the training cost of the visual localization network.

Moreover, if the training samples used are not suitable for training semantic segmentation tasks in the field of unmanned driving technology, it may also cause the trained visual positioning network to be unable to accurately output pose data, thus giving unmanned equipment adversely affect the decision-making process

Images