SCI-OBS: Detection and Classification of People on the Beach and in the Surf Zone

Michael Reinwald , Ketron Mitchell-Wynne, Zane D'Amico, Mercedes Stringer, Benjamin Freeston

Understanding how many people use beaches and surf zones is vital for safety, tourism, and coastal management. Traditional manual headcounts are time-consuming and inconsistent, while existing automated systems often fail in complex conditions due to glare, waves obstructing the view, and the small apparent size of distant subjects. We present SCI-OBS, an automated multi-class detector deployed on the Surfline Coastal Intelligence (SCI) pipeline's 233-camera real-time GPU inference fleet (part of a broader ~1,200-camera archival network); this paper validates its performance on the people-detection task - detecting and counting people both on the beach and in the water - with the non-person classes (boats, jetskis, and lifeguard flags) deferred to later work. SCI-OBS is the Ultralytics YOLO26-L end-to-end detector fine-tuned on Surfline's proprietary coastal-camera dataset: tens of thousands of manually-labelled frames containing several hundred thousand labels drawn from the full diversity of Surfline's globally-distributed camera network. No coastal-specific architectural modifications were made. We validate the system on 1,400 frames from 14 representative camera sites across Australia, the United Kingdom, and the United States (64,130 ground-truth labels), benchmarking against six out-of-the-box baselines: YOLO26-L and RT-DETR-L from COCO-pretrained weights, each evaluated with and without Slicing Aided Hyper Inference (SAHI), plus MMDetection with SAHI and AWS Rekognition. SCI-OBS achieves the highest F1-score (84.0%) and recall (75.6%) of all evaluated systems while holding precision at 94.6%, and recovers on average 79.9% of the ground-truth count (mean absolute error of 9.57 individuals per frame). With each baseline evaluated at its best available operating point, none matches this precision-recall balance: the high-precision baselines reach their precision only by missing far more subjects (recovering as little as 12%), and the same architecture without coastal training reaches only 55.1% F1 (79.5% with slicing-aided inference). The gap is therefore one of training data, not architecture - closing it would require a comparable coastal-labelled corpus - and, decisively, even a detector that closed it would still emit only an undifferentiated "person" box. SCI-OBS instead resolves each detected person into a fine-grained recreation type - surfers by state, beachgoers by posture, stand-up paddleboarders - that general-purpose detectors structurally cannot produce, classifying on-sand versus in-water use with 94.4% accuracy on the same benchmark and the dominant recreation types reliably. This is what turns an undifferentiated headcount into the typed, on-sand-versus-in-water usage data that coastal-safety and management applications require.