ABSTRACT
In smart cities, a common problem is the parking spots classification
into empty and occupied. It may seem simple, but a large number
of Deep Learning approaches rely on CNNs (Convolutional Neural
Networks). These solutions are commonly expensive, demanding
high computational power and specialized hardware to run properly,
making them unsuitable for large-scale deployments, such as
in smart cities. In this work, we propose two lightweight CNN architectures,
built upon existing solutions by enhancing their efficiency
and robustness. We used a cross-dataset scenario, where a model
is trained and validated in two datasets and tested in another, applying
three robust state-of-the-art datasets: PKLot, CNRPark-EXT
and PLds. This process improves generalization across different
contexts and sets a more realistic scenario when compared to real
urban environments. Also, we compared our models to state-ofthe-
art networks, such as MobileNetV3 Large and Small to ensure
consistency and validate the results with well-explored models
in the literature. Our results showed that our models, with up to
34× and 88× fewer parameters than the MobileNetV3 Large, reach
less than 2% lower accuracy when compared to the MobileNetV3
networks. Furthermore, by using grayscale images, the results were
slightly better and also decreased processing and storage costs.