SmartSport: crowd counting meets large language models for smart facility management

Abstract Efficient management and optimal allocation of urban public sports facilities represents a critical challenge in modernizing public sports service governance. Current approaches to facility utilization assessment predominantly depend on manual counting and questionnaire surveys, which are constrained by low efficiency, limited coverage, and poor real-time responsiveness. Meanwhile, conventional management methods lack the capability to extract actionable insights from heterogeneous data sources, hindering intelligent decision-making. To address these limitations, this study introduces SmartSport, an intelligent management framework for public sports facilities that synergizes computer vision with large language models. The framework consists of two core modules. The CrowdVision module constructs a crowd counting network based on a lightweight visual state space model, leveraging selective state space mechanisms for global context modeling, cross-scan modules for spatial dependency capture, and multi-level feature pyramids for enhanced multi-scale perception. A point query mechanism is further incorporated to enable precise counting alongside crowd localization. The LLM-Advisor module employs large language models as the analytical engine, integrating crowd time-series data with geographic information, demographic distributions, and other contextual data in a structured manner. Through carefully designed prompt engineering, the model performs systematic analysis across three dimensions: usage pattern recognition, supply-demand gap diagnosis, and optimization recommendation generation, ultimately producing comprehensive management reports with problem attribution and actionable configuration suggestions. Experimental results on our constructed dataset demonstrate that the CrowdVision module achieves a counting accuracy of 93.8%, while recommendations generated by the LLM-Advisor module received a practicality score of 4.2 out of 5.0 based on blind evaluation by domain experts.