HOG-Pedestrian-Detector

This repository contains the code for a MATLAB implementation of a basic HOG + SVM pedestrian detector form my Computer Science Master thesis

Disclaimer

If you are going to use this code, please read the LICENCE and keep in mind that I PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY OF ANY KIND.

I partially adapted this code-base to newer versions of MATLAB but is very likely you find discrepancies in how some MATLAB functions work. I am in general happy to help understanding the project if you ask nicely but since the implementation of the project is now several years old and MATLAB has evolved, some functions might behave differently and I won't be updating the project continuously nor answering about how to update the code to newer MATLAB versions.

Requirements

MATLAB >= R2017b
libsvm >= 3.22

Installation

Please refer to MATLAB and libsvm documentation to install.

Data

Run

The project was developed on a Windows machine and now being resurrected on a Linux one, so you should be good in any platform as long as you can run MATLAB.

Setting the environment

Add the libs directory and all sub-directories to MATLABs path. Either through the command window or through the GUI.
Make sure libsvm is visible to MATLAB. If you are not sure if your installation of libsvm went alright, you can check with:

which -all svmtrain

Which should show something like:

~/HOG-Pedestrian-Detector/libs/libsvm-3.22/matlab/svmtrain.mexa64
/usr/local/MATLAB/R2017b/toolbox/stats/stats/svmtrain.m                 % Shadowed

There are several entry points to the project, but here the two main ones are shown:

Train

Assuming there's a models directory where trained models will be saved and that the positive and negative images can be found in dataset/Test/pos and dataset/Test/neg respectively. Train an SVM model named test

model = train_svm("test", ["./models", "dataset/Train/pos" "dataset/Train/neg"]);

Eval

To evaluate the just trained model:

test_svm(model.test, ["dataset/Test/pos" "dataset/Test/neg"]);

Note test_svm expects model.<model-given-name-to-train-function>...

PCA versions of train / test

[model, Ureduce] = train_svm_PCA("test_pca", ["./models", "dataset/Train/pos" "dataset/Train/neg"]);
test_svm_PCA(model.test_pca, Ureduce, ["dataset/Test/pos", "dataset/Test/neg"]);

Known issues & contributions

Old MATLAB version used to concatenate strings by enclosing them between squared brackets but doesn't look like valid any longer. In that case you should use the strcat function. For example when constructing paths, so:

path = ['folder', 'filename', '.extension']  % this is wrong!
path = strcat('folder', 'filename', '.extension')  % this is right!

If you enjoyed this repository and find things that are not working any longer, you are very welcome to open a PR with fixes and I'll happily introduce them.

baucheng / hog-pedestrian-detector Goto Github PK