Benchmarking Visual Detection Tiers for Event-Centric Edge Intelligence in Maritime Disaster Response

Rapid maritime incident response requires visual recognition near the point of observation, but practical deployment must balance detection accuracy, inference latency, hardware-conversion constraints, and limited communication capacity. This study benchmarks the visual-compute component of an Event-Centric Edge Intelligence (ECEI) architecture on two maritime edge tiers: a Jetson AGX Orin 64 GB Developer Kit and a reComputer AI R2000-12 based on Raspberry Pi 5. Three tasks were selected to represent distinct maritime event categories: Human Detection for person-overboard response, six-class Ship Detection for vessel monitoring and collision awareness, and eleven-class Marine Trash Detection for floating-debris and pollution monitoring. YOLO26, YOLO11, and YOLOv8 were evaluated across 54 model–dataset–device configurations. Repeated on-device validation used 100 deterministically selected images per task and 20 measured repetitions for each checkpoint, yielding 108,000 timed predictions. On the paired S/N subset, the median reComputer/AGX mean-latency ratio was 4.40×; mean latency ranged from 19.84–57.09 ms on AGX and 68.89–172.11 ms on reComputer. The highest AGX mAP50–95 values were 0.331, 0.813, and 0.639 for Human, Ship, and Marine Trash Detection, respectively; the highest reComputer values were 0.318, 0.806, and 0.558. Across the evaluated event-message configuration, event-candidate JSON bytes were 98.20% lower than raw JPEG bytes. These results support deployment-aware visual-tier selection and an auditable detector-to-event interface for maritime edge intelligence.