Memory Enabled Segmentation of Terrain for Traversability based Reactive Navigation

Abstract

This article presents a novel 2D traversability image estimation for local reactive navigation, that attributes the fusion of a novel Convolutional Neural Network (CNN) for coarse semantic segmentation on terrain roughness, with surface geometric normals. The proposed segmentation model consists of a U-Net based Encoder-Decoder architecture with a MobileNet V3 Large backbone for real-time performance. At the bottom layer, the bottleneck block commonly found in a U-Net has been enhanced with an Atrous Spatial Pyramid Pooling (ASPP) block. In addition, a SEResNet based decoder instead of the classical stacked convolution blocks of U-Net has been implemented, while a concatenation layer has been added at the output. Moreover, the development of a novel memory module to dynamically update the semantic segmentation image based on certainty heat maps is also shown. The efficacy of the proposed scheme has been evaluated in real-life environments such as indoors, outdoors and subterranean (SubT) environments on a Pioneer 3AT mobile robot.