FalconEye: Efficient Centroid-Based Object Detection for Maritime High Altitude UAV Images on Embedded Devices

Maritime search and rescue operations increasingly utilize Unmanned Aerial Vehicle based object detection; however, existing methods often struggle to balance detection accuracy and the computational limitations of embedded hardware. To address this challenge, we introduce Falcon-Eye, a lightweight neural architecture optimized for detecting small objects in high-altitude maritime imagery. Our approach challenges the trend of using complex architec-tures by adopting a streamlined single-branch design with minimal downsampling. Further, we propose the Within-Bounds Metric, which demonstrates that precise bounding box predictions may not be necessary for search and rescue applications. We also introduce a new comprehensive dataset specifically designed for maritime search and rescue scenarios. Evaluation on this dataset and an existing benchmark shows that Falcon-Eye achieves competitive performance, exceeding 90% mean Average Precision, while reducing the number of parameters by 95% compared to existing methods. The model also provides significant latency improvements across platforms, making it especially suitable for resource-constrained Unmanned Aerial Vehicle deployments. Our findings suggest that it is possible to develop simpler and more efficient solutions that can rival current state-of-the-art approaches in maritime object detection by carefully considering practical requirements.