Volltext-Downloads (blau) und Frontdoor-Views (grau)
  • search hit 3 of 18
Back to Result List

Data simulation in deep learning-based human recognition

  • Human recognition is an important part of perception systems, such as those used in autonomous vehicles or robots. These systems often use deep neural networks for this purpose, which rely on large amounts of data that ideally cover various situations, movements, visual appearances, and interactions. However, obtaining such data is typically complex and expensive. In addition to raw data, labels are required to create training data for supervised learning. Thus, manual annotation of bounding boxes, keypoints, orientations, or actions performed is frequently necessary. This work addresses whether the laborious acquisition and creation of data can be simplified through targeted simulation. If data are generated in a simulation, information such as positions, dimensions, orientations, surfaces, and occlusions are already known, and appropriate labels can be generated automatically. A key question is whether deep neural networks, trained with simulated data, can be applied to real data. This work explores the use of simulated training data using examples from the field of pedestrian detection for autonomous vehicles. On the one hand, it is shown how existing systems can be improved by targeted retraining with simulation data, for example to better recognize corner cases. On the other hand, the work focuses on the generation of data that hardly or not occur at all in real standard datasets. It will be demonstrated how training data can be generated by targeted acquisition and combination of motion data and 3D models, which contain finely graded action labels to recognize even complex pedestrian situations. Through the diverse annotation data that simulations provide, it becomes possible to train deep neural networks for a wide variety of tasks with one dataset. In this work, such simulated data is used to train a novel deep multitask network that brings together diverse, previously mostly independently considered but related, tasks such as 2D and 3D human pose recognition and body and orientation estimation.

Download full text files

Export metadata

Additional Services

Search Google Scholar

Statistics

frontdoor_oas
Metadaten
URN:urn:nbn:de:bsz:rt2-opus4-45224
DOI:https://doi.org/10.15496/publikation-81046
Publisher:Universität Tübingen
Place of publication:Tübingen
Referee:Cristóbal CurioORCiD, Andreas Schilling
Referee of HS Reutlingen:Curio, Cristóbal
Document Type:Doctoral Thesis
Language:English
Publication year:2023
Date of final exam:2023/04/04
Tag:Fußgänger; Klassifikation; deep learning; maschinelles Sehen; neuronales Netz
Page Number:205
Dissertation note:Dissertation, Universität Tübingen, 2023
DDC classes:004 Informatik
Open access?:Ja
Licence (German):License Logo  Open Access