A Transfer Learning Approach for Real-Time Parking Availability Classification Using MobileNetV2 and Efficient Net

Abstract

It is hard for computers to tell if parking is available because of things like different parking lot layouts, different types of vehicles, and photos that are low quality or grainy. We propose a transfer learning method based on TensorFlow Hub to create a large-scale multi-class image classification system for determining whether a parking space is available. To improve the model's performance, pre-trained architectures such as MobileNetV2 and Efficient Net are utilised to create rich feature representations from images of parking lots. We utilise a publicly available dataset with annotated images of parking lots and cars for training and testing. To improve the quality and reliability of the data, various preprocessing methods are employed, including image scaling, normalisation, and augmentation. We use categorical cross-entropy loss to improve the model, and we evaluate it using Top-1 and Top-5 accuracy metrics. The results demonstrate that this method is effective and can be applied in real-time parking availability detection, smart city solutions, and automated parking management systems. We also examine problems, including class imbalance, vehicles blocking each other, and misclassification, as parking layouts are identical. This work demonstrates how transfer learning can be beneficial for classifying large numbers of images, particularly for use on edge devices and in real-time applications that track parking availability.