Static 3D Map Reconstruction based on Image Semantic Segmentation

Abstract

In this paper we propose a novel approach to remove dynamic objects from the reconstructed 3D map. We use a pixel-wise, fully convolutional neural network (FCN) based image semantic segmentation method to detect the moving objects and coherently remove them. Both 2D and 3D filters are employed to deal with the stubborn unwanted pixels to preserve as much information as possible. Experiment shows our approach can effectively build static maps.