Sparse Convolutional Neural Networks for Point-Cloud-Based Object Detection in Autonomous Driving

Institut
Lehrstuhl für Fahrzeugtechnik
Typ
Masterarbeit /
Inhalt
experimentell / theoretisch /  
Beschreibung

Successfully mastering the autonomous driving task depends highly on an accurate representation and understanding of the environment. To achieve such a detailed knowledge of the surrounding, current object detection algorithms use not just camera but also lidar or radar data. However, processing point cloud data is challenging due to its sparse, unordered data structure with a variable number of points and its high dimensionality. Therefore, specialized network structures are used to process this kind of data.

The objective of this thesis is the implementation of a spares convolutional neural network for object detection on automotive radar data. The model should be built upon the current state of the art and extend the idea of sparse convolutional neural networks to the radar domain. The adaption of sparse convolution operations should increase the performance of point-cloud-based network architectures and bring the success of image-based object detection to the point-cloud domain.

The first step of this project consists of a literature research on the current stat of the art in sparse convolutional neural networks and object detection on radar point cloud data. In the second step, a sparse convolutional neural network for object detection on radar point cloud data should be built upon the current state of the art and applied to different automotive dataset. Finally, the results of the network architecture should be compared to the current state of the art and an outlook on future network architectures should be given.

Work packages

  • Literature research on sparse convolutional neural networks
  • Implementation of a sparse convolutional neural network for object detection
  • Comparison of the results of different network architectures on multiple datasets
  • Deduction of an outlook on future improvements on the model design

Requirements

Voraussetzungen
  • Programming experience in Python
  • Involved working attitude
  • Ideally experience with Docker
  • Ideally experience in machine learning

The thesis can be written in German or English language.

Tags
FTM Studienarbeit, FTM IVS, FTM Fent, FTM Informatik
Möglicher Beginn
sofort
Kontakt
Felix Fent, M. Sc.
Raum: MW3508
Tel.: +49 89 289 15347
fentftm.mw.tum.de
Ausschreibung