Abstract
Urban growth and the increasing number of vehicles have intensified
the demand for efficient parking management solutions. In this
context, machine learning-based image monitoring systems have
gained prominence due to their low cost and ease of installation
compared to traditional methods, such as physical sensors. These
systems achieve an average accuracy of 95% in cross-validation
scenarios using well-known datasets like PKLot and CNRPark-EXT.
However, despite the availability of extensive datasets, challenges
remain regarding the accessibility and diversity of training data.
This is especially critical when aiming to improve the accuracy of
generalist models or specialize them for specific scenarios, where
each application requires a substantial effort to collect, segment, and
label new images for optimal performance. This study proposes the
use of synthetic images, generated with the Unity 5 graphics engine
and the Unity Perception package, to complement or replace real
data in training parking classification models. A synthetic image
generation protocol was developed to reduce costs compared to the
collection, segmentation, and labeling of real images. The images
generated through this protocol are referred to as low-fidelity due
to their lower quality and reduced capacity to simulate specific environments.
Using MobileNetV3 and transfer learning, experiments
were conducted in three scenarios: total replacement of real data,
supplementation of diverse datasets, and specialization for specific
scenarios. The results showed that synthetic images could improve
model generalization by up to 2% in datasets with limited real data
(e.g., CNRPark-EXT). However, synthetic images alone could not
fully replace real data due to their limited fidelity in replicating
real-world conditions, reinforcing the need for combinations with
real data or more realistic synthetic data for better results.