Deep Learning based Dynamic-to-Static Image Translation for Accurate Localization of Autonomous Driving Vehicles

Simultaneous localization and mapping is widely used for real-time localization of autonomous vehicles when GPS signals are not available. However, the presence of dynamic objects in the city can lead to the introduction of incorrect feature matching pairs, resulting in degradation of localization accuracy. To tackle this challenge, we focus on dynamic-to-static image translation, aiming to recover static images from dynamic scenes and reduce the negative effect of dynamic elements. We achieve this by proposing a semantic-guided dynamic-to-static image translation model. It recovers static semantic segmentation maps from the dynamic image and then leverages the encoded semantic feature maps to guide the generation of static images, making the model more focused on semantic information. Specifically, in this paper, we first segment the images semantically and then fuse the up-sampling layer of the semantic segmentation network with the down-sampling layer of the dynamic-static translation model. Experiments show that the model can better perform the task of dynamic scene translation and help unmanned vehicles to achieve accurate navigation and positioning based on vSLAM.