Using convolutional neural networks for car parking occupancy monitoring

This report presents an approach to car parking occupancy detection that uses techniques of deep learning analysis on the video stream produced by surveillance cameras. In order to validate the developed algorithms, we have also produced a dataset consisting of training set and test set. We have made the dataset publicly available to the scientific community, to develop and compare new techniques of car parking occupation monitoring based on video stream anal- ysis. The dataset contains pictures of parking slots taken from different viewpoints and with different light conditions. It also contains occlusions and shadows that might disturb the classification of the parking slot status. Our experiments show that deep learning is very effective for the occupancy detection task. It is very robust to light conditions changes, presence of shadows, and partial occlu- sions. In addition, it provides good results also when tests are performed using video stream captured from a viewpoint different than the viewpoint used for training.