car-classifier-research

This repo combines research documentation, source files and results from two final year university research subjects

1. COMP4110 Computer Vision
   • Authors: myself and Dylan Barnett:
   • This course tasked us to conduct some computer vision based research within a group, for which we chose to focus on training object detectors using feature extraction
2. COMP4120 Directed Studies
   • Author: myself
   • This course allowed for any directed study for which i chose to extend the previous car detection research.

Project Goals:

Train object detector to -

• detect cars ahead, from road view
  Train with -
  
• Cascaded Haar-features
• Cascaded Haar-features & Viola Jones Framework
• Cascaded Local Binary Patterns (LBP)
  Test with -
  
• color (normal) vs grayscale
• clean (normal) vs noisy
• light (normal) vs dark (night, tunnels) images
• (normal) vs low powered (raspPi)

The research is currently underway.

Setup:

The process of setting up openCV on Windows machine is found here

We started with a clone of https://github.com/mrnugget/opencv-haar-classifier-training And followed the following method:

Create list of samples from images:

find ./negative_images -iname "*.jpg" > negatives.txt

find ./positive_images -iname "*.jpg" > positives.txt

create many positive samples from provided ones

$ perl src/createsamples.pl positives.txt negatives.txt samples 1500 "opencv_createsamples -bgcolor 0 -bgthresh 0 -maxxangle 0.01 -maxyangle 0.01 maxzangle 0.01 -maxidev 40 -w 20 -h 20"

nb. warnings may come up when running the perl script, since i had to change some files from png to jpg. This is because of [stricter libpng rules] (http://tex.stackexchange.com/questions/125612/warning-pdflatex-libpng-warning-iccp-known-incorrect-srgb-profile) but it still works with them.

go to samples dir and run merge script

python mergevec.py -v ../samples -o samples.vec

run training

opencv_traincascade -data classifier -vec samples/samples.vec -bg negatives.txt -numStages 20 -minHitRate 0.999 -maxFalseAlarmRate 0.5 -numPos 1000 -numNeg 13 -w 300 -h 175 -mode BASIC -precalcValBufSize 512

nb. [numPos needs to be somewhat less than the number of postive samples, based on background similarity, numStages, minHitRate] (http://stackoverflow.com/questions/10863560/haar-training-opencv-assertion-failed)

General rule is: (numPos = 0.9 * total samples) will work most of the time. Or 0.8 x numberOfElements in vec file to not risk failing.

Useful links:

[openCV docs on Cascade classifiers] (http://docs.opencv.org/doc/user_guide/ug_traincascade.html)

Descriptors

Coding Robin training classifier

Naotoshi Seo training classifier

Rotations when generating samples

[Ubuntu manual] (http://manpages.ubuntu.com/manpages/precise/man1/opencv_createsamples.1.html)