摘要
为了提高即时定位与地图构建(SLAM)系统的容错能力,在经典图像生成网络Pix2Pix的基础上,逐步添加深度估计网络和深度信息的输入、基于STN网络的图像重建损失以及基于图像修复网络的图像补全损失3个方面的改进.结合双目图像的耦合关系,通过挖掘和融合多种信息,增大了信息的利用率,提高了模型的图像生成效果.提出将生成对抗网络(GAN)技术与SLAM容错场景相结合,直接实现了感知端的容错.在KITTI和Cityscapes数据集上进行实验,验证了改进模型的有效性.将模型生成的图像用于双目视觉系统的重建,验证了容错思想的可行性.
The classical Pix2Pix network was modified in order to promote the capacity of fault tolerance of simultaneous localization and mapping (SLAM) system. The network was gradually added to depth estimation network and its depth information, image reconstruction loss based on STN network and image inpainting loss based on image inpainting network. Information was mined based on the coupling of stereo images and merged to utilize information usage and promote model performance. Then generative adversarial net (GAN) and SLAM were combined, and the fault tolerance in the sensing level was directly realized. Experiments were performed on KITTI and Cityscapes dataset in order to prove the effectiveness of the improvement. The generated images and original images were both fed as inputs of stereo SLAM system. Results showed that the fault tolerance idea was approachable.
作者
王凯
岳泊暄
傅骏伟
梁军
WANG Kai;YUE Bo-xuan;FU Jun-wei;LIANG Jun(College of Control Science and Engineering, Zhejiang University, Hangzhou 310058, China)
出处
《浙江大学学报（工学版）》
EI
CAS
CSCD
北大核心
2019年第1期115-125,共11页
Journal of Zhejiang University：Engineering Science
基金
国家自然科学基金资助项目(U1664264
U1509203)
