A simulated car-park environment for the evaluation of video-based on-site parking guidance systems

Developing image-processing algorithms based on machine learning is a challenging problem concerning the huge amount of thoroughly annotated data needed. The internet provides many already tagged images for basic classification problems like vegetables or different cars, but not for more narrow problems. In order to extend and evaluate the previously presented parking guidance system from our previous work, in this paper, we propose a simulation system based on Unreal Engine 4. We developed an artificial camera which implements all features of a real camera, e.g., lens distortion, motion blur etc. to export video data from the simulated environment. This data is then compared to real-world video footage by using our classification module that distinguishes occupied and free parking lots. We reached a classification rate between 92.28 % and 99.72 % depending on the parking rows' distance using DoG-features and a support vector machine.