To address challenges such as the failure and inefficiency of crowd evacuation guidance in low-visibility environments, this study constructs an evacuation model that incorporates the micro-interaction mechanism between Unmanned Aerial Vehicles (UAVs) and pedestrians. Integrating the social force model with the optimal reciprocal collision avoidance algorithm, the model meticulously characterizes the dynamic “perception-decision-motion” feedback loop for both pedestrians and UAVs. Furthermore, it formally defines three guidance interaction modes: auditory, visual, and communication, and formulates guidance strategies based on these. System simulation results demonstrate that multi-modal UAV guidance universally and significantly enhances evacuation performance. In all test scenarios, overall evacuation efficiency (measured by average travel time) is superior under UAV guidance compared to random exploration. Even the visual mode, which exhibited the lowest performance, achieved a reduction in average travel time of approximately 9.99% under the most unfavorable conditions. Conversely, the communication mode demonstrated optimal performance, reducing average travel time by up to 26.36%. Sensitivity analysis reveals that evacuation performance is highly sensitive to the information transmission radius and information acceptance rate, but relatively insensitive to information granularity. Additionally, the study uncovers interaction effects between interaction modes and flight strategies, identifying the combination of the communication mode with a fixed-route strategy as the most effective. A critical finding indicates that while the auditory mode improves overall evacuation efficiency, it may compromise individual safety (measured by maximum travel time), highlighting an intrinsic trade-off between collective efficiency and individual safety. This research provides a quantitative analytical framework and optimization principles for the design of intelligent evacuation systems in low-visibility conditions.