The characterisation and simulation of 3D vision sensors for measurement optimisation

The use of 3D Vision is becoming increasingly common in a range of industrial applications including part identification, reverse engineering, quality control and inspection. To facilitate this increased usage, especially in autonomous applications such as free-form assembly and robotic metrology, the capability to deploy a sensor to the optimum pose for a measurement task is essential to reduce cycle times and increase measurement quality. Doing so requires knowledge of the 3D sensor capabilities on a material specific basis, as the optical properties of a surface, object shape, pose and even the measurement itself have severe implications for the data quality. This need is not reflected in the current state of sensor haracterisation standards which commonly utilise optically compliant artefacts and therefore can not inform the user of a 3D sensor the realistic expected performance on non-ideal objects. This thesis presents a method of scoring candidate viewpoints for their ability to perform geometric measurements on an object of arbitrary surface finish. This is achieved by first defining a technology independent, empirical sensor characterisation method which implements a novel variant of the commonly used point density point cloud quality metric, which is normalised to isolate the effect of surface finish on sensor performance, as well as the more conventional assessment of point standard deviation. The characterisation method generates a set of performance maps for a sensor per material which are a function of distance and surface orientation. A sensor simulation incorporates these performance maps to estimate the statistical properties of a point cloud on objects with arbitrary shape and surface finish, providing the sensor has been characterised on the material in question. A framework for scoring measurement specific candidate viewpoints is presented in the context of the geometric inspection of four artefacts with different surface finish but identical geometry. Views are scored on their ability to perform each measurement based on a novel view score metric, which incorporates the expected point density, noise and occlusion of measurement dependent model features. The simulation is able to score the views reliably on all four surface finishes tested, which range from ideal matt white to highly polished aluminium. In 93% of measurements, a set of optimal or nearly optimal views is correctly selected.