摘要
为促进自动驾驶技术的发展,采用基于多模态图像深度学习的汽车虚拟驾驶环境生成方法,同时生成包含多个物理场景的多模态图像。利用部分共享的隐空间构建编码器和生成器,采用域不变特性的感知损失,通过Cityscapes和Comma2k19数据集进行对比实验,采取多样性评价指标进行评价。结果表明:采用多模态深度学习生成的虚拟驾驶环境图像具有高真实性和多样性,对于快速构建自动驾驶虚拟仿真平台具有重要意义。
To facilitate the development of autonomous driving technology,a generation method for automobiles virtual driving environments based on deep learning of multimodal images is used to generate multimodal images which contain multiple physical scenes simultaneously.The encoder and generator are constructed using a partially shared hidden space,and the perceptual loss of domain-invariant properties is used,and the comparison experiments of Cityscapes and Comma2k19 datasets are carried out.The diversity evaluation metrics are adopted for evaluation.The results show that the virtual driving environment images generated by multimodal deep learning is high realism and diversity,which are important for the rapid construction of the virtual simulation platform for autonomous driving.
作者
张书生
祝雪峰
叶乾
ZHANG Shusheng;ZHU Xuefeng;YE Qian(School of Automobile Engineering,Dalian University of Technology,Dalian 116024,Liaoning,China;Ningbo Research Institute of Dalian University of Technology,Ningbo 315000,Zhejiang,China)
出处
《计算机辅助工程》
2023年第4期23-28,共6页
Computer Aided Engineering
基金
国家重点研发计划(2021YFB3300601)
中央高校基本科研专项资金(DUT22QN241)。
关键词
虚拟驾驶环境
深度学习
多模态
自动驾驶
图像生成
仿真平台
virtual driving environment
deep learning
multimodal
autonomous driving
image generation
simulation platform