Robust human instance segmentation in a challenging forest environment

Abstract

A key requirement of safe robotic operation in environments where a mobile robot shares a space with people is that it is able to recognise and avoid situations where it may collide with a human. This is particularly important in uncontrolled environments or rugged terrain where navigation options dynamically change and are usually constrained in some way. We propose a sensor system that will form part of such a safe navigation system that is able to robustly identify and localise people in a densely vegetated forest environment even when they are occluded by vegetation. It accomplished this by combining sensor data from thermal and colour cameras, and generating segmentation masks using a modified version of Mask R-CNN.