Spectral graph features for improved object level place recognition in dynamic environments

Abstract

Robotics and automation are rapidly growing fields in research, consumer markets, and industry. In recent years, we have seen the emergence of autonomous mobile robots. To operate autonomously, these mobile robots must be able to sense and understand their environment. This process is known as simultaneous localization and mapping (SLAM). During the SLAM process, a robot uses external sensors to build a map of its surroundings while simultaneously localizing itself within that map. One particular challenge in SLAM systems is odometry drift. As a robot maps its environment, errors in sensor calibration and readings will result in the accumulation of pose errors. Over time, this produces skewed and misaligned global maps. To address this issue of odometry drift, SLAM systems utilize place recognition systems. Place recognition allows a robot to recognize previously visited locations in its map. Once a place is recognized, a loop closure event is triggered to realign the global map. This Master’s thesis addresses the problem of conducting place recognition in scenes with dynamic objects. I introduce a method for object-level place recognition called Spectral Graph Place Recognition (SGPR). This method matches point cloud object-instances between scenes using geometrically encoded spectral features. These spectral features are calculated through a novel Minimally Connected Adaptive Radius algorithm and a Geometric Laplacian formulation to produce robust object-level matching for place recognition with dynamic objects. Using my research findings, I have created an open-source place recognition framework and an evaluation pipeline for others to utilize.