Vývoj a zhodnocení konceptu pro rozšíření dat pro trénování neuronových sítí pro semantickou segmentaci LiDAR-Pointclouds

Abstract

An accurate environment perception is a key requirement for automated vehicles. To help ful-fill this requirement, a promising approach is the semantic segmentation of LiDAR-pointcloudsusing neural networks. The acquisition of large amounts of annotated data needed for trainingthese networks is challenging due to the high cost of manually labeling pointclouds.In this work, data augmentation is proposed to obtain a high number of annotated LiDAR-pointclouds without the need of manual labeling. Two augmentation strategies are presented,the first one being the creation of semi-artificial samples and the second one the application oflabel-preserving transformations. Semi-artificial samples are created by automatically extract-ing objects from a controlled environment and inserting them into scenes where these objectsdo not occur.The results suggest that semi-artificial pointclouds can be successfully used as a supplementor as a substitution of real data. An intersection over union of 31.7 % can be achieved on a dataset of real world scenes when only semi-artificial pointclouds are used for training. Additionally,the recall on the real world data set can be increased if the semi-artificial data set is furtheraugmented with label-preserving transformations.

An accurate environment perception is a key requirement for automated vehicles. To help ful-fill this requirement, a promising approach is the semantic segmentation of LiDAR-pointcloudsusing neural networks. The acquisition of large amounts of annotated data needed for trainingthese networks is challenging due to the high cost of manually labeling pointclouds.In this work, data augmentation is proposed to obtain a high number of annotated LiDAR-pointclouds without the need of manual labeling. Two augmentation strategies are presented,the first one being the creation of semi-artificial samples and the second one the application oflabel-preserving transformations. Semi-artificial samples are created by automatically extract-ing objects from a controlled environment and inserting them into scenes where these objectsdo not occur.The results suggest that semi-artificial pointclouds can be successfully used as a supplementor as a substitution of real data. An intersection over union of 31.7 % can be achieved on a dataset of real world scenes when only semi-artificial pointclouds are used for training. Additionally,the recall on the real world data set can be increased if the semi-artificial data set is furtheraugmented with label-preserving transformations.