Volltext-Downloads (blau) und Frontdoor-Views (grau)

Analyse von Unsicherheiten künstlicher neuronaler Netze und Integration in die Objektverfolgung

  • Over the last few years, the development of assistance systems for motor vehicles has shifted from comfort functions to control tasks. Increasingly, these control tasks are also being transferred to semi-autonomous systems. One safety-critical aspect is the correct and reliable observation of the immediate environment by the vehicle. These observations can be used, among other things, to set up models for tracking objects. Due to recent research, topics such as uncertainties for object detections and the calibration of artificial neural networks are now emerging. The goal of this work is to investigate the possibility of processing positional uncertainties of a detector in a multiple object tracking approach and the e� ects on the tracking of objects. Additionally, the calibration of the used detector will be evaluated and corrected if necessary. The e� ects of the calibration on the tracking results will also be investigated in this context. After an investigation of the procedure used to generate the position uncertainties of the detector, a connection to the multiple object tracking was made and an approach to process the uncertainties based on a Kalman filter was developed. The confidence of the detections was also remodeled. For this purpose, the confidence was interpreted as the existence probability and processed using a Bayes Filter to reflect the existence of the tracks. In addition, appropriate calibration methods for the position uncertainties and confidence were selected and incorporated into the tracking procedure. The validation of the presented approaches was performed on a data set for driving situations. The evaluation of the results showed that a processing of the position uncertainties generated by a detector is feasible in the presented tracking approach. The interpretation of the confidence as existence probability leads to good results. Calibration of the confidence further improves the results. However, the calibration of the position uncertainties led to worse results. Further inves-tigation of other calibration methods for the position uncertainties is needed. Keywords: Multiple Object Tracking, Kalman Filter, Neural Network Calibration

Download full text files

Export metadata

Additional Services

Search Google Scholar


Author:Raphael Sebastian Burkert
Document Type:Master's Thesis
Year of Completion:2022
Date of final exam:2022/03/08
Release Date:2022/04/22
Page Number:56
Institutes:Fachbereich 1 - Institut Informatik
DDC class:300 Sozialwissenschaften / 330 Wirtschaft
Licence (German):License LogoNo Creative Commons