Datenbestand vom 13. Juni 2019

Warenkorb Datenschutzhinweis Dissertationsdruck Dissertationsverlag Institutsreihen     Preisrechner

WICHTIGER HINWEIS
DER VERLAG IST IN DER ZEIT VOM 12.06.2019 BIS 23.06.2019 AUSCHLIESSLICH PER EMAIL ERREICHBAR.

aktualisiert am 13. Juni 2019

ISBN 9783843909518

Euro 96,00 inkl. 7% MwSt


978-3-8439-0951-8, Reihe Informatik

Martin Rapus
Component-based Pedestrian Recognition for Advanced Driver Assistance Systems

302 Seiten, Dissertation Friedrich-Schiller-Universität Jena (2012), Softcover, A5

Zusammenfassung / Abstract

This book addresses the challenging task of pedestrian recognition in an urban environment for automotive applications. For automatic collision avoidance, it is necessary to estimate the 3d position and motion of pedestrians, even if they are partially occluded. To deal with this problem, a component-based recognition approach is presented, which has considerable advantages compared to other approaches in case of occlusions. Contributions are provided on several levels. Both, high performance and real-time capabilities are relevant and important criteria that are considered in this book.

For robust detection, the single-frame recognition approach integrates local object shape features. To select the most important features from a redundant set of shape features for optimal classification, different algorithms are developed. The component model is used for the detection and tracking of pedestrians. A human gait model is proposed for differentiating between plausible and impossible pedestrian movements during tracking. For robust tracking, an interest points model is integrated, which dynamically adapts to the posture of the tracked pedestrian. An important objective is to prevent the false activation of safety measures, e.g., brakes or steering. Therefore, an efficient hidden-visible state model is introduced to reduce incorrectly initialized tracks.

Other contributions concern the achievement of real-time capabilities. Therefore, a novel method based on optical flow is introduced to rapidly restrict the classification and tracking to important image locations. Furthermore, the detection process is speeded up by incorporating a cascade of support vector machine classifiers.

Several experiments are conducted to evaluate each level of the proposed pedestrian recognition framework. The corresponding datasets are taken from a moving vehicle in a complex urban environment. Single-frame classification results show a similar or even better performance compared to state of the art approaches. The performance of the hidden-visible state model highlights that the framework substantially improves the recognition statistics. An image set containing partially occluded pedestrians is used to evaluate the performance of the framework with respect to partial occlusions. In case of partial occlusions, a significant performance gain of the component-based approach compared to existing approaches can be achieved.