Hugo Delasnerie

This paper explores the feasibility of training a neural network on digitally created data with the end goal of the neural network being able to perform object recognition on LiDAR point clouds. LiDAR is a form of 3D scanning which can produce a real time 3D model of whatever environment it is pointed at. A LiDAR emits large amounts of laser beams in a cone or sphere configuration. The beams reflect back to the LIDAR  and depending on how long it takes for the LiDAR to get the beams back, it can determine the distance of a surface. From all these points of contact, it generates a point cloud, which is a 3D representation of the space. This 3D model constantly updates as the LIDAR moves around in the environment. Finding an easy yet reliable way to generate large amounts of training data for a neural network to isolate and identify individual objects in a point cloud is essential. The goal of this research is to prove that creating digitally simulated point clouds from digital 3D environments for use in neural network training is a feasible method for training a LiDAR neural network. The 3D environments were made and animated in Blender. The LiDAR simulation of these environments was done in Isaac Sim, and the neural network is programmed in Python. So far, the LiDAR simulation of 3D environments has been very successful, producing large amounts of data and accurate point clouds. Even in a complex 3D environment, the LiDAR simulator was quick and accurate, producing detailed point clouds with perfectly realistic amounts of noise and very accurate velocity data. If we can prove the feasibility of using simulated point clouds for neural network training, then the process by which those networks can be trained will be greatly shortened.