An Adaptive Across Time–Space-Domain Perception Method for Rescue Robots Based on DN-CycleGAN and MR-Xception Transfer Learning

Recently, rescue robots have been utilized in scenarios such as earthquakes, chemical explosions, etc. The challenges for these robots lie in terrain awareness and target recognition by machine vision. The rescue environments are complex with harsh weather changes and challenging terrain, leading to differences in the temporal and spatial domains between the visual training and test data. Traditional crossdomain sample set collection methods are time-consuming and heavily rely on large sample sizes. Furthermore, obtaining a comprehensive training sample set under extreme weather conditions including various time domains (e.g., ice, snow) and spatial domains (e.g., fog, poison gas) is difficult. This paper proposes an across time-space environment adaptive perception method for rescue robots based on an improved CycleGAN termed DN-CycleGAN. With support of a terrain-diversified sample synthesis method based on style transfer, it can enhance the transition layer and discriminator of the cycle network generator. Compared with the traditional CycleGAN, the DN-CycleGAN in this paper has a 5.26% higher PSNR score and a 9.27% higher SSIM, and the DN-CycleGAN obtains a more realistic image to a certain extent, with color characteristics closer to the target style image. In this paper, MR-Xception, a terrain classification algorithm based on Xception transfer learning is proposed, which can cope with the low accuracy of traditional deep neural networks in terrain recognition and training time, which can deal with low accuracy of traditional deep neural networks in terrain recognition and training time. This method enables the diverse cross-domain recognition of small sample terrains and enhance the terrain recognition accuracy to 99.54%. It also ensures the transfer of the fundamental terrain style across various time-space domains. Moreover, this method enables the cross-domain migration of terrain dataset for the task of robot’s terrain recognition across different domains. Finally, it facilitates rapid visual perception and self-adaptation of robots in cross-domain tasks within unstructured environments.