AI-Driven Innovations in Smart Parking and Graph Neural Networks: A Survey

Abstract

The rapid progress of Intelligent transportation systems (ITS) is desperately needed to reduce the challenges of urbanism; today, a major challenge is to "cruising for parking" that causes around 30% of urban traffic, systemic density, and environmental impacts such as carbon emissions. In this paper, a comprehensive review of the evolution of urban mobility from fundamental hardware sensors and statistical baselines to high-rise learning and large-language model architecture (LLM) is presented. Despite significant gains in predictive accuracy, current deep learning paradigms exhibit substantial research gaps, notably the "black-box" nature of model inference, which lacks an intuitive mapping from multi-modal inputs to predicted results. Methodologically, existing spatio-temporal models are frequently constrained by fixed distance-based adjacency matrices, which fail to capture the "functional synchronization" between urban functional areas that are functionally similar but geographically non-adjacent. Furthermore, the transition to large-scale urban grids is currently hindered by unsustainable training overhead and a significant domain gap between natural language and structured traffic data. The scope of this review encompasses visual occupancy detection backbones, hierarchical graph-based predictors, and the emerging generative AI frameworks for explainable forecasting. We conclude that while the integration of Graph Neural Networks (GNNs) and recurrent units has improved non-linear modeling, the LLM paradigm offers superior representation for few-shot learning and context-aware reasoning. However, critical limitations regarding inference latency and the scarcity of seasonally diverse benchmarks remain unresolved. Future research must prioritize adaptive, efficient, and transparent frameworks capable of modeling complex dependencies across heterogeneous urban environments.