Swin-GNN: A Spatial Reasoning Framework for Intelligent Multi-Level Parking Occupancy Detection

Abstract

Vision-based parking occupancy detection has progressed from hand-crafted image descriptors and classical classifiers to deep learning models such as convolutional neural networks, vision transformers and graph neural networks. Despite this progress, most methods classify each parking space in isolation and rarely exploit the spatial relationships between neighbouring spaces. The limitation is most pronounced in multi-level indoor parking structures, where lighting differences, structural obstructions, vehicle density, perspective distortion and level-to-level dependence all degrade per-space detection. This paper introduces Swin-GNN, a Swin Transformer–Graph Neural Network framework for parking occupancy detection. A Swin-Small backbone extracts hierarchical visual features from parking-space image crops; each space is then represented as a graph node with spatial adjacency encoded as graph edges, and a three-layer Graph Attention Network propagates occupancy evidence between neighbouring spaces before a lightweight multilayer perceptron produces the final decision. Hyper-parameters are selected with a genetic algorithm. The framework is validated on the PKLot benchmark using a curated hard-condition test set that combines five real-world degradation factors (illumination variation, occlusion, compression artefacts, perspective distortion and temporal class imbalance), and PKLot’s lack of indoor multi-level scenes is explicitly acknowledged as a limitation. Swin-GNN achieves 96.59 % accuracy, an F1-score of 0.965 and an AUC of 0.992 with 45 ms per-image latency, 22.2 images/s throughput and 99.2 % service-level compliance. The architecture, training protocol, hyper-parameter optimisation and a factor-wise degradation analysis are reported in full.