Intelligent Urban Parking Recommendation System using Spatiotemporal Graph Neural Network, Large Language Model and Vision Assistant

The growing imbalance between urban parking demand and available capacity leads to excessive cruising, traffic congestion, and unnecessary emissions, which collectively degrade urban traffic efficiency. To address this challenge, this paper proposes multimodal Spatiotemporal Graph Neural Network, Large Language Models and Vision Algorithm (STGNN-LLMaVA), framework for intelligent urban parking recommendation that jointly models visual perception, semantic context, and dynamic spatiotemporal dependencies. Parking-slot occupancy is inferred from surveillance images using You Only Look Once version 12 (YOLOv12). In parallel, the Large Language model and Vision Assistant (LLMaVA) generate compact semantic and temporal descriptions of the scene, capturing factors such as congestion, visibility, and surrounding activity. These visual and language-derived features are embedded into parking-node representations and processed by a GraphKAN&ndashTemporal Transformer backbone. In this backbone, a Kolmogorov&ndashArnold Network models nonlinear spatial interactions, while a causal Temporal Transformer captures evolving availability patterns to produce top-k parking recommendations. Experiments conducted on four real-world parking datasets demonstrate that STGNN-LLMaVA consistently outperforms strong Graph Neural Network (GNN)- and Large Language Model (LLM)-based baselines, achieving improvements of up to 24.7% in HitRate@10, 10.5% in NDCG@10, and 9.8% in MRR@10. These results indicate that integrating vision-based occupancy estimation, language-driven contextual reasoning, and spatiotemporal graph learning provides an effective, scalable, and data-efficient solution for sustainable smart parking management.