On-street parking space localization with deep learning using low-quality images from public cameras

Abstract

The increasing demand for new services in cities leads to challenges that need an intelligent, holistic approach (smart cities). A crucial aspect of smart city planning is effectively managing public on-street parking spaces, influencing overall urban mobility and environmental sustainability. However, detecting these spaces is complex, often requiring costly cameras or sensors, and the variability in parking space sizes, depending on the vehicles parked, adds to the difficulty. Existing city cameras for traffic monitoring could be a solution, but their low image quality and frequent movements (taking images from different angles) make accurate detection challenging. We propose a novel method for locating on-street parking spaces using low-quality images from non-static public traffic cameras. This approach is dataset-independent, applicable to various cities, and employs deep-learning models pre-trained for tasks like vehicle detection, repurposing them for the novel task of identifying on-street public parking spaces. This method avoids specific retraining and intensive manual labeling. Tested in Malaga, Spain, the pipeline includes Extraction (sourcing images from Internet traffic cameras), Matching (recognizing common features between reference and new images for detecting camera movements), Preprocessing (comparing different denoising and image-enhancing techniques for improving model inference), Detection (using models like YOLOv8 and Detectron2 for vehicle detection), and Postprocessing (transforming perspectives to estimate real-world parking space coordinates and sizes). Experimental results demonstrate that our proposal achieves accurate parking space detection even in extreme light conditions and camera movements, providing a valuable new tool for parking management and urban planning.