Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Image based Wheel Detection using Random Forest Classification

Hultström, Karin LU (2013) In Master's Theses in Mathematical Sciences 2013:E7 FMA820 20131
Mathematics (Faculty of Engineering)
Abstract
The aim of this master thesis is to detect and recognise wheels in images by means of image analysis. This could later on serve as a foundation for a safer vehicle counting and classification method than those currently in use that requires personnel to cross the lanes on installation.

The general layout of the classification system consists of five stages: multi-scale transformation, window extractor, pre-processing, classification and cluster analysis. In order to obtain the training and testing data for evaluation and construction of the system, images that illustrate moving cars on a road are acquired. From these, several positive and negative windows are extracted that visualizes wheels and non-wheels. For the classification... (More)
The aim of this master thesis is to detect and recognise wheels in images by means of image analysis. This could later on serve as a foundation for a safer vehicle counting and classification method than those currently in use that requires personnel to cross the lanes on installation.

The general layout of the classification system consists of five stages: multi-scale transformation, window extractor, pre-processing, classification and cluster analysis. In order to obtain the training and testing data for evaluation and construction of the system, images that illustrate moving cars on a road are acquired. From these, several positive and negative windows are extracted that visualizes wheels and non-wheels. For the classification stage, the learning algorithm used is Random Forest. Moreover, with the Random Forest as the foundation, two different concepts were introduced to further improve the predictions. These are referred to as bootstrap configuration and cascading classification.

The results are evaluated be means of Receiver Operating Characteristics and contingency tables. In this master thesis, the final system produces a satisfying result based on the false positive rate and true positive rate. For future development, the amount of examples in the training data could be increased in order to gain more knowledge in the teaching of the classifier. Furthermore, an optimization of the program could lead to faster execution time, which is a requirement if this system is to operate in real-time. To conclude, the system produces a satisfying result for wheel detection that can be used as a foundation when constructing a general system for vehicle counting and classification. (Less)
Please use this url to cite or link to this publication:
author
Hultström, Karin LU
supervisor
organization
alternative title
Bildbaserad Hjuldetektion med hjälp av Random Forest Klassificering
course
FMA820 20131
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Classification, Detection, Random Forest, Decision Trees, Receiver Operating Characteristic, Vehicle counting, Wheel, Local Binary Pattern, Bootstrap Aggregating, Cascading
publication/series
Master's Theses in Mathematical Sciences 2013:E7
report number
LUTFMA-3240-2013
ISSN
1404-6342
other publication id
2013:E7
language
English
id
3457767
date added to LUP
2013-04-25 17:09:45
date last changed
2013-04-25 17:09:45
@misc{3457767,
  abstract     = {The aim of this master thesis is to detect and recognise wheels in images by means of image analysis. This could later on serve as a foundation for a safer vehicle counting and classification method than those currently in use that requires personnel to cross the lanes on installation. 

The general layout of the classification system consists of five stages: multi-scale transformation, window extractor, pre-processing, classification and cluster analysis. In order to obtain the training and testing data for evaluation and construction of the system, images that illustrate moving cars on a road are acquired. From these, several positive and negative windows are extracted that visualizes wheels and non-wheels. For the classification stage, the learning algorithm used is Random Forest. Moreover, with the Random Forest as the foundation, two different concepts were introduced to further improve the predictions. These are referred to as bootstrap configuration and cascading classification.

The results are evaluated be means of Receiver Operating Characteristics and contingency tables. In this master thesis, the final system produces a satisfying result based on the false positive rate and true positive rate. For future development, the amount of examples in the training data could be increased in order to gain more knowledge in the teaching of the classifier. Furthermore, an optimization of the program could lead to faster execution time, which is a requirement if this system is to operate in real-time. To conclude, the system produces a satisfying result for wheel detection that can be used as a foundation when constructing a general system for vehicle counting and classification.},
  author       = {Hultström, Karin},
  issn         = {1404-6342},
  keyword      = {Classification,Detection,Random Forest,Decision Trees,Receiver Operating Characteristic,Vehicle counting,Wheel,Local Binary Pattern,Bootstrap Aggregating,Cascading},
  language     = {eng},
  note         = {Student Paper},
  series       = {Master's Theses in Mathematical Sciences 2013:E7},
  title        = {Image based Wheel Detection using Random Forest Classification},
  year         = {2013},
}