Pedestrian Detection

Note: The code provided may not be compatible with python3. A migration to python3 is in the pipeline.

Dependencies

• OpenCV
• scikit-image pip install scikit-image==0.14.1
• scikkit-learn pip install scikit-learn==0.20.2

Running Detection

To test on images, simply run, python detectmulti.py -i <path to image>

For example, python detectmulti.py -i sample_images/pedestrian.jpg

For more options run, python detectmulti.py -h

Following are some examples,

Training

This code is meant for Training a Pedestrian Detector using INRIA Person Dataset.
Note: If the above link doesn't work, then the above dataset can be downloaded from here.

Download, Uncompress and place it in the root of this repository.

Then just run:

sudo sh fixpng.sh # To fix the broken png files in this dataset.

Note: The above step is necessary to fix the broken pngs in the dataset, without running this the classifer won't be correctly trained.

To train just run:

python train.py --pos <path to positive images> --neg <path to negative images>

For INRIA dataset, this would be probably,

python train.py --pos INRIAPerson/train_64x128_H96/pos --neg INRIAPerson/train_64x128_H96/neg

After training, two new files would be created, namely, person.pkl and person_final.pkl. The former is the pre eliminary detector and the latter is the improved (hard negatively mined) detector.

Note on Training:

Training can use high amounts of memory, so be sure to have a swap space in case of RAM overflows. Also, memory consumption can be reduced by decreasing the maximum number of hard negative windows to be mined. This is defined by the MAX_HARD_NEGATIVES global variable in train.py.

Testing

After successful training just run:

python test.py --pos <path to positive images> --neg <path to negative images>

For INRIA dataset, this would be probably,

python test.py --pos INRIAPerson/test_64x128_H96/pos --neg INRIAPerson/test_64x128_H96/neg

This would print True Positives, True Negatives, False Positives, False Negatives, Precision, Recall and F1 Score.