Point cloud semantic map construction method based on deep learning and laser radar

Abstract

The invention relates to the technical field of intelligent semantic maps, more particularly to a point cloud semantic map construction method based on deep learning and a laser radar. The method comprises: step one, constructing a deep convolutional neural network; step two, carrying out laser radar point cloud preprocessing; step three, training a deep convolutional neural network model; step four, inputting a laser radar point cloud into the deep convolutional neural network model to obtain a tag of each point in the laser radar point cloud being one including a semantic annotation; step five, carrying out real-time point cloud map construction by using the laser radar point cloud including the semantic annotation so as to obtain a point cloud semantic map; and step six, after construction of the point cloud semantic map, carrying out global optimization and correction on the semantic information in the point cloud map by using voting based on a sliding window.

Description

technical field [0001] The present invention relates to the technical field of intelligent semantic maps, and more specifically, to a point cloud semantic map construction method based on deep learning and laser radar. Background technique [0002] Lidar point cloud maps play an important role in the perception and motion planning of autonomous robots (including but not limited to: autonomous vehicles, robots, unmanned aerial vehicles, etc.). There are many methods for generating laser point cloud maps based on lidar. LiDAR point cloud maps typically consist of large-scale, discrete points. Therefore, the lidar point cloud map lacking semantic information is more difficult for human eyes to identify, which is not conducive to manual post-processing. At the same time, due to the lack of semantic information in the map, it is difficult for engineers to formulate more complex motion planning for robots. At present, there have been researches on semantic segmentation of objec...

AI Technical Summary

Problems solved by technology

In the process of using multiple sensors to construct static road networks and dynamic obstacle maps, it is necessary to perform a series of sensor calibration operations to fuse each sensor in advance. This calibration process is relatively complicated and has errors.

At the same time, GPS sensors cannot be used indoors; high-definition satellite images are difficult to obtain, so the process of building static road networks and obstacle maps with multiple sensors at the same time is cumbersome and not suitable for indoor scenarios and promotion

In the prior art, the semantic markup of the map is related to people, and each map needs to be manually marked separately, which is a large workload for the construction of large-scale semantic maps

Images