SAFETYLIT WEEKLY UPDATE

We compile citations and summaries of about 400 new articles every week.
RSS Feed

HELP: Tutorials | FAQ
CONTACT US: Contact info

Search Results

Journal Article

Citation

Iurgel U, Park SB, Schauland S. VDI Berichte 2006; 2006(1960): 625-638.

Affiliation

Delphi Delco Electronics Europe GmbH, Wuppertal

Copyright

(Copyright © 2006, VDI Verlag)

DOI

unavailable

PMID

unavailable

Abstract

The European Commission has promoted the bisection of the number of accidents by 2010 in their eSafety program [1]. Especially the protection of pedestrians is supposed to be enhanced essentially by the European legislation. In spring 2003 the European Commission adopted a directive for pedestrian protection, defining limits of injury heaviness of the various body zones of pedestrians accidents with cars. One of the various ways to decrease pedestrian accident involvement is the development of pedestrian detection systems able to detect a pedestrian in front of the vehicle and warn the driver or actively intervene before an accident happens. This paper presents the feature extraction and classification modules of a vision-based pedestrian detection system. Similar to most systems for pedestrian detection, our system is based on a combination of features, namely overcomplete wavelet features and a combination of symmetry and edge density features. Other features used are Legendre Moments and Edge Orientation Histogram (EOH). This new constellation of feature representations leads to excellent test results in terms of run-time and classification performance, since classifiers can be cascaded in order to lower overall computation costs. As we will show, especially the combination of features is favorable. Support Vector Machines are used as classifier, since they are easy to handle and to implement (compared to Neural Networks). Based on theoretical facts, fast learning algorithms are available, and they have been proved to achieve very good experimental results. Classifiers have to be evaluated in order to measure and predict their performance. Evaluation must not be based on subjective impressions, but has to make use of a statistical analysis. But statistical measures alone do not necessarily reflect the performance of the classifier under real operating conditions. It is also important to correctly design the classifier training and its evaluation. These aspects will be discussed in detail. Specifically, we use an evaluation measure not simply based on accuracy, but on a more advanced statistical measure based on precision and recall: the F-measure allows to prioritize the classification results during training to avoid false positives (FP) at the cost of false negatives (FN). Choosing the right statistical measure for training and testing has been disregarded by most research teams in the past while playing an important role in achieving good results for the final system.

NEW SEARCH


All SafetyLit records are available for automatic download to Zotero & Mendeley
Print