University of Twente Student Theses


Point cloud based semantic segmentation for catenary systems using deep learning : Compressibility of a PointNet++ network

Rutgers, Nils (2022) Point cloud based semantic segmentation for catenary systems using deep learning : Compressibility of a PointNet++ network.

[img] PDF
Abstract:The current signalling system used to manage train traffic is outdated. The entire railroad system is divided into sections. As only one train is allowed to be in a specific section, the track is used inefficiently. ProRail is collaborating with StruktonRail to modernise the signalling system. To get an overview of the current catenary systems, a digital twin needs to be created that digitally resembles the physical situation. A deep neural network was trained using point clouds, in this case, a PointNet++ network, to identify catenary arches. To deploy such a model on a microcontroller, it needs to be compressed. There are numerous algorithms to compress a computer vision model, however, there is no method available to compress a network trained on point clouds. This research aims to investigate the feasibility of applying available compression methods on a PointNet++ network while preserving performance. The research will follow the CRISP-DMME cycle with a focus on the technical understanding and realisation phases. Literature research is done to determine a suitable method for compressing a PointNet++ network. This method leverages pruning, quantization and Huffman encoding to compress the size of a model. A digits classification and a PointNet++ model were compressed using this method. By comparing the compression results in terms of weight loss, size reduction and change in accuracy, the effectiveness of the compression method is evaluated. It is found that compression without performance loss is possible for computer vision networks. However, when compressing a PointNet++ using conventional compression methods, the accuracy will drop drastically. On the other hand, the models show similarities in terms of weight reduction and quantization. There are similarities between the compression results of both models. This indicates that compression without performance loss might be possible. However, further research needs to be done on the similarities between image and point cloud trained models on theoretical and software levels.
Item Type:Essay (Bachelor)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Creative Technology BSc (50447)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page