Improving a Background Model for Tracking and Classification of Objects in LiDAR 3D Point Clouds

Abstract: This thesis studied methods of improving a background model for a data processing pipeline of LiDAR point clouds. For this, two main approaches were evaluated. The first was to implement and compare three different models for detecting ground in a point cloud. These were based on more classical modeling approaches. The second was to use Deep Learning for semantic segmentation of point clouds and to use this information in a background filtering model with the hope of achieving better filtering of dynamic background. These methods were combined in a pipeline as an example of a possible application. The performance of the ground models was primarily evaluated based on their ability of classifying points as ground or non-ground. However, well performing ground models have further uses. Of the three models studied, the Hybrid model achieved most promising results. For semantic segmentation, RandLA-NET was used for its ability to process large scale point clouds at high speeds. Variations of the network was trained on simulated data for which all networks achieved similar good performance for classes ground, vegetation and other. When testing domain transfer to point clouds produced by a Real Physical LiDAR, mixed results were achieved with variations on a per-point-cloud basis. On a lot of instances however, very promising results could be seen. A background subtraction model based on a 3D Density Static Filter was extended to include semantic information from the neural network. For this filter, voxels classified as vegetation and their neighbours, heavily filtered out points. This was to avoid issues of false detections caused by wind. The model was tested on parts of two LiDAR recordings and compared to the standard filter. Based on this, the extended model was found to better filter out vegetation in windy conditions.