The search for a parking space in urban areas is often time-consuming and nerve-racking. Efficient car park guidance systems could support drivers in their search for an available parking space. Video-based systems are a reasonably priced alternative to systems employing other sensor types and their camera input can be used for various tasks within the system.
Current systems detecting vacant parking spaces are either very expensive due to their hardware requirements or do not provide a detailed occupancy map. While several sensor types feature individual parking space surveillance, their installation and maintenance costs are relatively high. The system developed in this research group has minimal hardware requirements, which makes it less expensive and easy to install. At the same time, our video-based approach offers flexibility regarding information usage and site of operation.
The system consists of one or more wide-angle lens cameras in combination with a standard desktop computer. The cameras can be optionally equipped with a microcomputer each to calculate metadata before sending it through the network, thus complying with possible data protection guidelines. Each camera can monitor up to 36 parking lots, depending on its position.
Once set up and calibrated, the system uses different image features and machine learning algorithms to determine whether or not an individual parking space is occupied. This classification is acquired in real time, i.e. on 5 frames per second. The resulting information on vacant parking spaces can be requested via smartphone app.
We trained and tested a number of SVM- and kNN-classifiers on HoG- and color features, e.g. (Tschentscher, 2015), of occupied and free parking space snippets. After the training phase the system was able to classify parking spaces with an accuracy of 99.8 % employing temporal filtering.
Synthetic Data Usage
In order to meet the requirements of machine learning algorithms regarding the necessary amount of data, a simulated environment has been developed to create further image data for the evaluation of existing classifiers and the training of a new classifier for the parking space detection task. The simulated environment allows the automatic generation of large amounts of image data and ground-truth information as well as the reconstruction of special cases by full human interaction control. Variable image data was created this way, including different lighting and weather conditions. Some of our results can be seen in this video:
In order to validate the approach of using synthetic image data as training input for machine learning algorithms, we used the same classifiers as before (trained on traditional real-world images) and evaluated them on simulated video data with comparable results (Tschentscher, 2017).
Finally, we trained a new classifier on purely simulated image data and evaluated its performance on unseen natural video sequences (Horn, 2018). The results were similar to those produced with the previously described classifiers.
- Best Poster Paper Award at the 2017 IEEE Intelligent Vehicles Symposium (IV' 2017)
- News Portal der Ruhr-Universität Bochum: Simulierter Parkplatz im Nebel (11/2017)
- IEEE ITSS Newsletter: 2017 October Edition: Best Paper Awards (10/2017)
- Parken aktuell: Kamerabasierte Klassifizierung von Parkplatzbelegungen (04/2013)
Evaluation of Synthetic Video Data in Machine Learning Approaches for Parking Space Classification