Real-Time Sea Turtle Detection with YOLO Optimized for Edge Hardware

Abstract

Sea turtles are critically endangered species whose conservation depends on continuous monitoring of nesting beaches. On-foot patrols are labor-intensive and limited in coverage, while satellite telemetry is cost-prohibitive for fixed-site monitoring. This work presents an embedded real-time detection system for sea turtles based on YOLO object detection models running on a Raspberry Pi 5. A dataset of 11,096 images has been created, combining public sources with locally captured images and hard negative samples. Three models were trained and evaluated, namely: YOLO26n, YOLO11n, and YOLO11s, using internal and external validation protocols. The selected model (YOLO26n) achieves a false positive rate of 0.98% and external recall of 92.85%, exported via ONNX to the NCNN framework and executed in C++ for maximum throughput. The system captures video via a CSI-connected Pi Camera through the libcamera API, operates at 12.82 FPS with a mean latency of 76.2 ms, and delivers real-time alerts through an SSE notification pipeline to a web-based monitoring dashboard over a local Wi-Fi hotspot, with all components automatically started at boot via systemd services, enabling autonomous, internet-independent operation suitable for coastal preservation areas.