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, and what they are doing there, is vital for safety, tourism and coastal management. Camera-based studies of coastal recreation have to date recovered attendance: a headcount, sometimes divided into coarse zones. This paper reports the accuracy at which the same infrastructure resolves attendance into activity composition, using a multi-class detector deployed in Surfline Coastal Intelligence’s coastal monitoring pipeline. The detector is a one-stage convolutional architecture fine-tuned on 27,712 manually-labelled coastal-camera frames, validated on 1,400 frames recorded under clear visibility at 14 sites in Australia, the United Kingdom and the United States against five out-of-the-box baseline models. Each baseline is given its best achievable operating point on the validation data while the presented model is held at its production threshold, so the reported margins are lower bounds. The model attains the highest precision (94.8%), recall (78.5%) and F1-score (85.8%) of the systems evaluated, and its F1 exceeds every baseline’s at every confidence threshold; its margin over the strongest baseline is 3.0 F1 points, with a site-level bootstrap interval excluding zero. False detections make up 5.2% of its reported count, less than half the share carried by the two closest baselines, whose aggregate counts rely in part on false positives cancelling missed subjects. Each detected person is then resolved into a recreation type: on-sand and in-water use separate with 97.2% accuracy, and eight activity classes are typed at 83.6% support-weighted and 70.1% macro-averaged accuracy. Resolving a frame’s headcount into typed on-sand and in-water counts costs 0.73 subjects per frame over the untyped count of the same subjects. That breakdown is the form coastal-management and water-safety applications require, and no general-purpose person detector can produce it.