This is an implementation of the HOG + SVM approach for pedestrian detection first done by Dalal & Triggs in their paper Histograms of Oriented Gradients for Human Detection in 2005.
In this approach, we parse a window of fixed size over the whole image, and for each window we get a fixed number of HOG features which are then passed to a SVM to classify if a person is present or not.
- Get HOG features using opencv or skimage on images of size 64x128.
- Detect pedestrians in the image using SVM.
ETHZ
To create the positive samples we can used the annotation provided by them and negative samples were generated by sampling some random points in each image. Since, the number of pedestrians are generally less random sampling gives us negative samples.
We have tried 2 ways to detect pedetrians in the image.
(i) Load the pre-trained Linear SVM model from openCV.
(ii) Train a SVM with rbf kernel only on this dataset and see the results after NMS (Non-max suppression).
The first approach involves using a HoGDescriptor from openCV whoce parameters can be set either using variables or using a .xml file. Here we have used the hog.xml file to set the parameters for the HOG.
We set the SVM detector for this by using setSVMDetector() and then passing the HOGDescriptor_getDefaultPeopleDetector() from opencv into it. We put a threshold on the weight so that trees are not detected as pedestrians, which is quite a common error for a pre-trained model. Opencv_HOG contains the scripts for this.
In the second approach since there are not many negative samples given in the datatset hence choosing random windows and labelling them negative does work but still it is tough for the model to differentiate between the pedestrians and the side-walk. Hence, we decided to stick with the Linear SVM pre-trained on INRIA dataset. HOG_SVM_train conatins the notebook for training the SVM and testing.
The blue rectangles represent the ones not selected by non-max supression.
This is an example when a non-human object gets chosen as a pedestrian.
