Efficient marking method combining laser point cloud and images

Abstract

The invention provides an efficient marking method combining laser point cloud and an image. The method comprises: performing initial external parameter automatic calibration through the planar checkerboard target image data and the laser point cloud data; pre-annotation is realized through an automatic segmentation algorithm; image annotation information is further refined in combination with a small amount of manual intervention and calibration; the method comprises the following steps: determining a three-dimensional laser point cloud corresponding to an image marking object in a back projection mode, then obtaining an accurate three-dimensional point cloud of a to-be-marked target in a re-segmentation clustering and growth mode, and finally performing external parameter further optimization through the accurately matched three-dimensional point cloud and the image marking object. According to the efficient marking method combining the laser point cloud and the image, a large amountof manual intervention is not needed, the difficulty of laser point cloud marking is reduced, the marking efficiency is improved, and meanwhile the high marking precision is achieved. Not only can the point-by-point category information of the laser point cloud be obtained, but also novel labeling data such as image and laser point cloud object level joint labeling information can be obtained.

Description

technical field [0001] The invention relates to the technical field of automatic driving, in particular to an efficient labeling method combining laser point clouds and images. Background technique [0002] In the field of unmanned driving technology, sensor fusion, especially low-level fusion, has become one of the effective technical solutions in unmanned driving technology, by identifying and estimating the semantics of key elements such as vehicles, lane lines, pedestrians, and traffic signs in the surrounding environment and geometric information to assist unmanned vehicles to perceive the environment and plan driving routes. The current mainstream method is to use lidar and cameras to perceive the environment. Lidar can obtain accurate geometric information of the observed environment, and cameras can obtain For images with rich texture and color information, more accurate environmental information can be obtained through the fusion of the two. [0003] This technolog...

AI Technical Summary

Problems solved by technology

[0003] This technology needs to rely on the deep learning network in practical application. However, training the deep learning network model based on low-level fusion data requires massive laser point cloud and image data sets that have been aligned in time and space and jointly labeled. Such data sets are currently It has not yet been formed, so it is of great significance to study new and effective tools for joint labeling of laser point cloud and image data sets

Images