Skip to Main content Skip to Navigation
Theses

Méthodes à noyaux pour la détection de piétons

Abstract : A lot of research have been carried out in the field of pedestrian detection using images and computers. The main obstacle is related with the pedestrian himself, which could not be easily characterized, due to its high variability. In particular, we have a scale, pose and appearance variability.
To bring an issue to these variabilities, the pedestrian representation should be carefully chosen. In our case, we first use a graph based representation. In fact, a labeled graph has interesting properties to reduce particularly the scale and pose variability.
A second representation is based on histogramms of oriented gradient, which computes some local histogramms of an image. This method has a good generalization capacity against variability. To apply this method on infrared images in order to detect pedestrians, we have to design a function to extract and analyse some windows from this image.
The second part of the pattern recognition process is the classification. In our case we use the Support Vector Machine classifier, which is based on an particular function : the kernel function. The aim is to compute the inner product between data.
For the graph method, we have to design an inner product between graphs. The aim is to compare two graphs by using bag-of-paths. That is to say, we extract some paths from graphs and we compare paths between them. The final kernel is obtained by combining all comparisons between paths.
We also studied different kinds of kernel functions more specific for the histogramm-based representation in order to improve the performance of the classifier when we are using the HOG descriptor.
Document type :
Theses
Complete list of metadatas

Cited literature [99 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-00375617
Contributor : Julien Comte <>
Submitted on : Wednesday, April 15, 2009 - 4:07:51 PM
Last modification on : Tuesday, February 5, 2019 - 11:44:22 AM
Long-term archiving on: : Thursday, June 10, 2010 - 8:34:13 PM

Identifiers

  • HAL Id : tel-00375617, version 1

Citation

Frédéric Suard. Méthodes à noyaux pour la détection de piétons. Informatique [cs]. INSA de Rouen, 2006. Français. ⟨tel-00375617⟩

Share

Metrics

Record views

359

Files downloads

1348