This study aims to explore the properties of point clouds of individual trees manually extracted from a large UAV LiDAR survey over a dense mixed forest. The dataset contains 192 trees of seven species, with an approximately even mix of deciduous and coniferous trees, which proves to be very limiting. Two experiments are described. The first is aimed at setting up an approach to help gain intuition about manual features used in classical machine learning on point clouds. Example visualizations are provided to demonstrate the effectiveness. The second is aimed at assessing the viability of approximating average tree shapes by estimating the kernel density of three-dimensional distributions of point coordinates. It shows relatively good overall accuracy in binary classification, but is unreliable in the multiclass setting and too slow. The dataset and the code are freely accessible through Kaggle.

The utilization of robots, for mechanical part assemblies, is a current practice. Existing methods and software permit to simulate tasks and program them off line using the perfect geometry of the parts. However, as the parts to be assembled are never perfect the adjustment time of these programs is often too long. It is indispensable to find models that can integrate these manufacturing uncertainties at the simulation level. The method proposed can replace the verification using manual functional gauges and can be integrated to a simulation software. In relation to the existing methods, it is especially suitable for the intended parts to be assembled and the datum reference frame (DRF) is defined at maximum material condition. This paper proposes a method for all situations of the DRF where the part is in conformance. It can be applied to inspection, as well as to tolerance analysis and synthesis. A kinematic model is used to represent the permissible motion of the DRF.