Research on critical criteria for crowd instability based on system entropy

During emergency evacuation, dense crowd aggregation in passages can trigger instability and stampede accidents, impairing evacuation and rescue effectiveness. This paper proposes an analytical method integrating computer vision and simulation to quantify crowd instability thresholds. Initially, accurate pedestrian detection is achieved using the YOLOv8n model trained on the CrowdHuman dataset, combined with the Deepsort algorithm to extract parameters (density, speed, and system entropy) from perspective-corrected accident scenes. Through analysis, a multi-dimensional instability criterion is derived. Video monitoring data is analyzed in simulation software (using AnyLogic state diagrams). Dynamic evaluation of multiple critical parameter thresholds is conducted through state diagram models, thereby enabling the technical integration mechanism between the two to be established. Analysis of incidents like the Itaewon stampede identifies critical thresholds: density (6.875 - 6.971 ped/m²), speed (0.177 - 0.179 m/s), and system entropy (555.796 - 582.194). Compared to single-density metrics, system entropy as a composite indicator more precisely captures multi-mechanism instability precursors, providing critical data support for early warning systems. Simulations indicate that passages with widths of 2.9 - 3.4 meters and lengths greater than or equal to 30 meters exhibit lower instability risks and higher pedestrian capacity. Sensitivity analysis reveals that the critical crowd size is more affected by passage width in flat areas and by length in sloped areas. The transition time from critical to safe pedestrian levels follows a linear distribution, with sloped passages exhibiting longer transition times and higher risks.