Autonomous drone swarms are able to offer flexible and rapid coverage for a variety of applications, including emergency logistics, disaster management, and urban monitoring. In order to coordinate a number of unmanned aerial vehicles (UAVs) at the same time, perception, communication, route planning and energy management must all work together rather than acting as separate units. Current methods generally improve these two areas independently, which leads to congestion on the wireless channel and poor tracking performance when the camera is moving quickly, serious drift during GPS outages in urban areas, and mission failures under degraded flight conditions. In this paper we propose an event-driven hierarchical AI decision-making system that integrates strategic routing with perception at the edge. At the higher level, Ant Colony Optimisation (ACO) is used to determine delivery routes that have the minimum makespan, and Adaptive Large Neighbourhood Search (ALNS) is applied to quickly re-route in the event of battery depletion or delivery cancellations. At the lower level, onboard edge perception is carried out using YOLOv8, confidence-based selective background compression, and Lucas–Kanade (LK) optical flow together with peer homography rectification (H_A->B) to eliminate camera drift. In order to ensure robust dead-reckoning and to prevent catastrophic divergence of position during urban GPS outages, a dual-layer TrajectoryLSTM recurrent network is trained to estimate three-dimensional flight momentum. The whole framework has been developed and tested in a high-fidelity Robot Operating System 2 (ROS 2 Humble) using a 3D simulation environment in Gazebo based on the VisDrone dataset. Our semantic compression reduces the transmission of background data by 96.89% (from 225.0 KB of raw frames to 7.0 KB) and cuts end-to-end communication latency from 413.6 ms down to 207.4 ms (a reduction of 49.9%) as empirically demonstrated. Moreover, ALNS reallocation reduces the mission makespan in the case of battery depletion from 327 seconds to 248 seconds (a 24.2% improvement), collaborative homography brings the tracking mIoU back to 0.717, and TrajectoryLSTM reduces the GPS-denied trajectory drift from 5.14 metres to 2.66 metres (a 48.35% reduction in error)