Parking Occupancy Detection

The implementation of Acharya, D., Yan, D., Khoshelham, K., (2018) Real-time image-based parking occupancy detection using deep learning. In this paper, the authors perform a pre-trained of Convolutional Neural Networks (VGGNet) and Support Vector Machine in order to detect parking occupancy.

The video for the related paper here. The authors' source code is found here here

Dataset: Parking Lot Database

This database contains 12,417 images (1280X720) captured from two different parking lots (parking1 and parking2) in sunny, cloudy and rainy days. The first parking lot has two different capture angles (parking1a and parking 1b).

The images are organised into three directories (parking1a, parking1b and parking2). Each directory contains three subdirectories for different weather conditions (cloudy, rainy and sunny). Inside of each subdirectory the images are organised by acquisition date.

More info about the dataset: https://web.inf.ufpr.br/vri/databases/parking-lot-database/

Code

• make_dataset.py - Build dataset

• model - Contain my baseline model and pre-trained model

  • baseline.py - My baseline
  • vgg.py - Apply the pre-trained VGG16

• train.py - Run training for model folder

• extract_feats.py - Use CNN model to extract features

• svm_clf_from_cnn_feats.py - Classifier

Usage

Preparation

1. Git clone https://github.com/anthng/Car-Parking-Occupancy-Detection.git
2. Download Dataset: link
3. Create folder: 'dataset', and then move the dataset zip into 'dataset' folder. Finally, unzip the dataset zip.
4. Open terminal in the root project folder: run pip install -r requirements.txt in your terminal.

Run

In your terminal:

• python make_dataset.py - build dataset
• python train.py - for training phase

For feature extraction by CNN, and SVM classifier

• python extract_feats.py - Use CNN model to extract features
• python svm_clf_from_cnn_feats.py - Run SVM classifier

My Result

I can not apply the pre-trained VGG16 model to extract features because of my hardware limitations. I used a MobileNet (is a light-weight model) instead of VGG16 as mentioned by the authors. The result was not gained as my expectation when applying MobileNet.

My Baseline

Pre-trained model

CNN - Feature Extraction and SVM Classifier

Acknowledgment

Source code for making dataset and training are built based on Real-time parking lot occupancy detection using Deep Learning