A Frequency-Domain Implementation of a Sliding-Window Traffic Sign Detector for Large Scale Panoramic Datasets

In large-scale automatic traffic sign surveying systems, the primary computational effort is concentrated at the traffic sign detection stage. This paper focuses on reducing the computational load of particularly the sliding window object detection algorithm which is employed for traffic sign detect...

Full description

Bibliographic Details
Main Authors: I.M. Creusen, L. Hazelhoff, P.H.N. De With
Format: Article
Language:English
Published: Copernicus Publications 2013-10-01
Series:ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Online Access:http://www.isprs-ann-photogramm-remote-sens-spatial-inf-sci.net/II-3-W2/7/2013/isprsannals-II-3-W2-7-2013.pdf
id doaj-96ad692d76e34718996f67d50f3bac7b
record_format Article
spelling doaj-96ad692d76e34718996f67d50f3bac7b2020-11-24T21:22:32ZengCopernicus PublicationsISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences2194-90422194-90502013-10-01II-3/W271110.5194/isprsannals-II-3-W2-7-2013A Frequency-Domain Implementation of a Sliding-Window Traffic Sign Detector for Large Scale Panoramic DatasetsI.M. Creusen0L. Hazelhoff1P.H.N. De With2Eindhoven University of Technology, Eindhoven, The NetherlandsEindhoven University of Technology, Eindhoven, The NetherlandsEindhoven University of Technology, Eindhoven, The NetherlandsIn large-scale automatic traffic sign surveying systems, the primary computational effort is concentrated at the traffic sign detection stage. This paper focuses on reducing the computational load of particularly the sliding window object detection algorithm which is employed for traffic sign detection. Sliding-window object detectors often use a linear SVM to classify the features in a window. In this case, the classification can be seen as a convolution of the feature maps with the SVM kernel. It is well known that convolution can be efficiently implemented in the frequency domain, for kernels larger than a certain size. We show that by careful reordering of sliding-window operations, most of the frequency-domain transformations can be eliminated, leading to a substantial increase in efficiency. Additionally, we suggest to use the overlap-add method to keep the memory use within reasonable bounds. This allows us to keep all the transformed kernels in memory, thereby eliminating even more domain transformations, and allows all scales in a multiscale pyramid to be processed using the same set of transformed kernels. For a typical sliding-window implementation, we have found that the detector execution performance improves with a factor of 5.3. As a bonus, many of the detector improvements from literature, e.g. chi-squared kernel approximations, sub-class splitting algorithms etc., can be more easily applied at a lower performance penalty because of an improved scalability.http://www.isprs-ann-photogramm-remote-sens-spatial-inf-sci.net/II-3-W2/7/2013/isprsannals-II-3-W2-7-2013.pdf
collection DOAJ
language English
format Article
sources DOAJ
author I.M. Creusen
L. Hazelhoff
P.H.N. De With
spellingShingle I.M. Creusen
L. Hazelhoff
P.H.N. De With
A Frequency-Domain Implementation of a Sliding-Window Traffic Sign Detector for Large Scale Panoramic Datasets
ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
author_facet I.M. Creusen
L. Hazelhoff
P.H.N. De With
author_sort I.M. Creusen
title A Frequency-Domain Implementation of a Sliding-Window Traffic Sign Detector for Large Scale Panoramic Datasets
title_short A Frequency-Domain Implementation of a Sliding-Window Traffic Sign Detector for Large Scale Panoramic Datasets
title_full A Frequency-Domain Implementation of a Sliding-Window Traffic Sign Detector for Large Scale Panoramic Datasets
title_fullStr A Frequency-Domain Implementation of a Sliding-Window Traffic Sign Detector for Large Scale Panoramic Datasets
title_full_unstemmed A Frequency-Domain Implementation of a Sliding-Window Traffic Sign Detector for Large Scale Panoramic Datasets
title_sort frequency-domain implementation of a sliding-window traffic sign detector for large scale panoramic datasets
publisher Copernicus Publications
series ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
issn 2194-9042
2194-9050
publishDate 2013-10-01
description In large-scale automatic traffic sign surveying systems, the primary computational effort is concentrated at the traffic sign detection stage. This paper focuses on reducing the computational load of particularly the sliding window object detection algorithm which is employed for traffic sign detection. Sliding-window object detectors often use a linear SVM to classify the features in a window. In this case, the classification can be seen as a convolution of the feature maps with the SVM kernel. It is well known that convolution can be efficiently implemented in the frequency domain, for kernels larger than a certain size. We show that by careful reordering of sliding-window operations, most of the frequency-domain transformations can be eliminated, leading to a substantial increase in efficiency. Additionally, we suggest to use the overlap-add method to keep the memory use within reasonable bounds. This allows us to keep all the transformed kernels in memory, thereby eliminating even more domain transformations, and allows all scales in a multiscale pyramid to be processed using the same set of transformed kernels. For a typical sliding-window implementation, we have found that the detector execution performance improves with a factor of 5.3. As a bonus, many of the detector improvements from literature, e.g. chi-squared kernel approximations, sub-class splitting algorithms etc., can be more easily applied at a lower performance penalty because of an improved scalability.
url http://www.isprs-ann-photogramm-remote-sens-spatial-inf-sci.net/II-3-W2/7/2013/isprsannals-II-3-W2-7-2013.pdf
work_keys_str_mv AT imcreusen afrequencydomainimplementationofaslidingwindowtrafficsigndetectorforlargescalepanoramicdatasets
AT lhazelhoff afrequencydomainimplementationofaslidingwindowtrafficsigndetectorforlargescalepanoramicdatasets
AT phndewith afrequencydomainimplementationofaslidingwindowtrafficsigndetectorforlargescalepanoramicdatasets
AT imcreusen frequencydomainimplementationofaslidingwindowtrafficsigndetectorforlargescalepanoramicdatasets
AT lhazelhoff frequencydomainimplementationofaslidingwindowtrafficsigndetectorforlargescalepanoramicdatasets
AT phndewith frequencydomainimplementationofaslidingwindowtrafficsigndetectorforlargescalepanoramicdatasets
_version_ 1725995471614246912