Mining is Australia's largest industry and generates over fty percent of the nation's export revenue. The on-going imperative to reduce the costs of mining motivates the automation of mining equipment. This thesis is focused on the automation of mining excavators that are capable of excavating 120 tonnes of ore in each dig cycle and then loading into a haul truck for transportation. To automate this process, the excavator needs to accurately know the relative position and orientation (or pose) of the trucks it is to load, and it needs to have condence in that pose.
This thesis addresses two problems that arise in the loading of trucks by automated excavators. The rst is the pose estimation problem: the determination of the pose of the truck relative to a frame of reference associated with the excavator. The second is the pose verication problem: given a pose of a truck, verify that the pose is within an allowed tolerance of error. The thesis uses planar scanning LiDAR range sensors xed to the machine house of a P&H 2100 BLE as the platform to address both problems. These sensors are orientated so that their scan planes map the surrounds of the machine as it
swings back and forth as part of its normal digging motion.
The thesis has made three major contributions to the truck pose problems that are identifed. The rst contribution is the development of a recursive method for estimating truck pose from planar range measurements within the framework of the Kalman lter. It utilizes a measurement prediction model that performs ray-casting against a geometric scene representation, which also facilitates automatic segmentation of data. The recursive estimator derives a new estimate each time a planar range scan measurement is made.
The second contribution is the development of a variable-structure extension to the recursive pose estimator, that accommodates the fact that not all of the degrees of freedom of the pose estimate are observable in planar range scan data. The variable-structure estimator identies the observable components of the truck pose from each new measurement. The observable and unobservable subspaces of the pose state space are decoupled by dening a fully-observable externally-equivalent system. The observable subspace can then be updated, before reforming the original truck pose state space. This allows the recursive estimator to accommodate the constraints imposed by the measurement process from excavator-mounted planar LiDAR sensors.
The third contribution is the development of a statistical hypothesis testing framework for pose verication, which seeks to verify a pose estimate is within an allowed tolerance region. The test makes use of density estimation techniques for each individual ray in the scanner measurement prole. These estimated univariate densities are combined to construct the multivariate density estimates required to compute the verication test statistic. The null hypothesis, that the pose estimate is sufficiently within the allowable tolerance of error, is rejected based on critical values of this statistic.