Determining and verifying object pose from LiDAR measurements to support the perception needs of an autonomous excavator

Abstract

This thesis is a study of how light detection and ranging, also known as LiDAR, can be used to meet perception requirements for field robot applications.LiDAR works by having an emitter produce a pulse of light along a known spatial direction. The distance or range to the world along that direction is determined by measuring the time it takes for reflections of the emitted light to be received at a collector. When this is done over multiple spatial directions, a cloud of points is produced that represent a sampling of the world. LiDAR is a mature technology and commercially available sensors are capable of producing point-clouds comprising up to 65,000 points over 360 degree fields of view, that are updated every 50 ms.This thesis is about how an autonomous machine can interpret meaning from these point-clouds in situations where the environment is complex and dynamic, where there is interaction with other agents leading to occlusion and where there is significant dust in the atmosphere.These conditions are systemic to open-cut mining environments, the domain where this thesis draws its motivation. In particular, the thesis uses as its primary example, the perception requirements for automation of a class of large mining excavators known as electric mining shovels. The work is specifically concerned with the problems of using point-clouds from machine mounted sensors to inform and verify knowledge of object position and orientation in the surrounds of these machines, for instance, the trucks being loaded.LiDAR is affected by dust and some see this a fatal flaw for mining robotics. The work starts with an exploration of the behaviour of a commercial LiDAR sensor in the presence of dust and finds that measurements are systematic and predictable. LiDAR sensors are found to exhibit four behaviours and it is shown how these behaviours can be understood from physics-based arguments. Several conclusions emerge, most notably where LiDAR measures dust, it does so to the leading edge of a dust-cloud rather than, say, as a random point within the cloud. This provides powerful insights into how better to use measurements from LiDAR when dust is present, and in particular suggests that dust can be treated like other forms of occlusion.A problem that must be solved to use LiDAR for machine perception is to establish from where each sensor views the world. This is sometimes called sensor registration and it amounts to finding where a sensor is positioned and how it is orientated. This is typically done by placing markers at known locations and computing the sensor registration from where they appear in the sensor data. For mining robot applications, the use of standard marker or artificial-feature-based approaches is infeasible. The thesis develops a method for sensor registration that overcomes this concern by utilizing the geometric structure of the terrain surrounding the autonomous vehicle platform. The method determines the i information content of registration parameters in measurements of the terrain, and update...

Real-Time 6-DOF Pose Estimation of Known Geometries in Point Cloud Data

“…This paper presents a solution that is straightforward to implement, does not require configuration, and meets the challenges of accuracy, robustness, and timeliness for a broad class of scenarios. The method is based on similar concepts presented in our previous work [ 1 , 2 ]. The issues of reliability, accuracy, and robustness are addressed by utilising reward-based metrics.…”