Automatic Parking Lot Classification
A Deep CNN Approach
Author: Andrews Sobral
First release: 25 September 2015
Last page update: 10 March 2017
Implementation: C++ and Python
Tools:
Caffe | Deep Learning Framework: http://caffe.berkeleyvision.org/
NVIDIA | DIGITS: https://developer.nvidia.com/digits
Image database:
PKLot dataset: http://web.inf.ufpr.br/vri/parking-lot-database
PKLot dataset paper: http://www.sciencedirect.com/science/article/pii/S0957417415001086
Summary:
The parking spaces were labeled manually, then a deep convolutional neural network (Deep CNN) tries to classify if each vehicle is present or not in each parking place. I have used a laptop computer to train the Deep CNN (only CPU mode), and the classification speed is very fast, i.e. less than 0.5 sec for 1 parking image with 28 parking places. It is important to note that each parking place is horizontal aligned and resized to 32x32 pixels. The neural network was trained considering three different situations: sunny, cloudy and rainy days (present in the PKLot dataset).
* Please see the PKLot dataset for a more detailed information about the parking scenes.
Related publications:
2016 Learning Approaches for Parking Lots
2016 QuickSpot - a video analytics solution for on-street vacant parking spot detection
2016 QuickSpot - a video analytics solution for on-street vacant parking spot detection
