A Semantic Segmentation Method for Street Scene Parsing for Autonomous Driving

Abstract

A semantic segmentation method for street scene parsing for automatic driving, comprising the following steps: constructing an image semantic segmentation network, the image semantic segmentation network is used to downsample an image to obtain an initial feature map, and a one-dimensional horizontal pooling is performed on the initial feature map. The first global feature map is obtained by performing a one-dimensional vertical pooling process on the initial feature map, and the first global feature map and the second global feature map are fused to generate an output image; Image training the image semantic segmentation network; using the trained image semantic segmentation network to perform semantic segmentation on the image to be processed. The invention utilizes the long and narrow pooling method of the global one-dimensional pooling mechanism, which can directly and effectively aggregate all the information in the horizontal and vertical directions, link a large amount of information to form effective context information, and make up for the The shortcomings of traditional rectangular pooling in aggregating long-distance context information.

Description

technical field [0001] The invention relates to the field of image semantic segmentation, in particular to a street scene parsing semantic segmentation method for automatic driving. Background technique [0002] Due to the increasing number of car users, the problems of road traffic congestion and safety accidents have become more and more serious. With the support of vehicle networking technology and artificial intelligence technology, autonomous driving technology can coordinate travel routes and planning time, thereby greatly improving travel efficiency and reducing energy consumption to a certain extent. For fast visual tasks such as automatic driving, the accuracy and efficiency of image semantic segmentation are very important, but the current semantic segmentation network cannot achieve a good balance between the two. [0003] At present, in order to improve the efficiency of semantic segmentation, a large number of lightweight network researches applied to real-time...

AI Technical Summary

Problems solved by technology

[0004] However, the existing network uses a two-dimensional square pooling operator to aggregate local area features to achieve feature pyramids and pyramid pooling modes with consistent size and proportion. This square pooling mode with a pooling aggregation range of a square area can only Aggregating object information in a local area cannot obtain effective long-distance context information

In addition, for some objects with irregular shapes and sizes, such as trees and utility poles, the two-dimensional square pooling operator will inevitably introduce irrelevant noise information, which will affect the network's overall understanding and judgment of features.

Images