IDEAL 2023

Conference: International Conference on Intelligent Data Engineering and Automated Learning 2023

Abstract:In large cities, a considerable amount of traffic congestion is caused by drivers actively looking for unoccupied parking spots, leading to reduced average velocities on the streets, reduced air quality, and an increase in sound pollution. A parking space map helps drivers to find vacant parking spaces more easily and helps minimising search time, fuel and energy consumption. We introduce a machine learning framework to detect occupied and empty parking spaces using airborne images. Our approach is based on three steps. First, we present novel ways to include observations from outside the visible spectrum (RGB) to delineate vegetation in urban environments. We remove roads on the images to minimise interference from moving vehicles and use a mask-RCNN to detect vacant parking spaces on the processed images. Third, we propose to use a logistic regression model (LRM) as a post-processing step in order to better filter true positives. From our experiments we conclude that Red-Green-NIR (RG-NIR) spectrum contains the most important information for training Faster-Region-based Convolutional Neural Network with Feature Pyramid Network (RCNN FPN) for vehicle detection. Our empirical studies on aerial images of Berlin, Germany, show that our framework can be employed successfully to create trustworthy parking space maps. Our best model obtain an accuracy of 97.7% and an F1-score of 0.645 for detecting both occupied and vacant on-street packing.

This is joint work with Bashini K. Mahaarachchi, Bodo Bookhagen, Vanja Doskoč and Tobias Friedrich.