Intelligent Parking Space Classification Under Hazy and Non-Hazy Conditions: An Efficient Deep Learning Solution

Abstract

A fundamental issue in managing parking effectively is optimizing the utilization of existing parking spaces. While advanced artificial techniques have demonstrated remarkable accuracy in classifying parking spots, optimizing their utilization remains a key concern. However, performance degrades during slight interference, obstructions, and diverse lighting conditions such as fog or haze. Most researchers have used deep learning approaches to classify parking spaces for non-hazy weather conditions only, often prioritizing model performance over training efficiency. Building on this, the main aim of the proposed work is to develop a model for classifying parking spaces under any weather conditions (hazy and non-hazy). A synthesized parking dataset is designed for hazy weather conditions. Light-dehazenet (LD-Net) is applied to counteract the effects of haze in the synthesized dataset. AlexNet is trained on the synthesized and PKLot datasets by applying transfer learning and preparing hazy and non-hazy feature vectors, respectively. A random forest is applied to select top-ranked features to avoid overfitting, remove noise from the features, and increase generalization capability. The selected features contribute to the input vector for classification using Multilayer-ELM (MLELM). The major breakthrough involves replacing the fully connected layer of AlexNet with MLELM to avoid lengthy backpropagation and reduce the training time. The experimental results of AlexNet-MLELM are compared with lightweight pre-trained CNN and existing state-of-the-art models. Empirical results suggest that the proposed model provides a viable approach for parking space classification in diverse weather conditions.