Visual feedback-driven active perception by drone swarms for proactive crowd anomaly capture

This study addresses the inefficiency and passivity of surveillance in detecting crowd anomalies across wide, dynamic environments using unmanned aerial vehicles.To address this, this paper proposes an active perception framework for drone swarms that is driven by real-time visual-semantic feedback.The framework couples a spatiotemporal graph attention network, which models crowd interactions and infers anomaly probabilities, with a cooperative multi-agent reinforcement learning decision-making module.This integration enables the swarm to dynamically and collaboratively optimise viewpoints based on live semantic cues.Evaluated on the VisDrone dataset, our approach achieves an anomaly capture rate of 89.7%, an average response delay of 1.9 seconds, an operational efficiency of 1.86 events per kilometre flown, and a low observation redundancy of 22.1%.These results demonstrate that embedding visual semantics into a closed perception-control loop significantly enhances the performance of proactive monitoring systems compared to existing baseline methods.