Artificial intelligence for beach monitoring: An experimental study of beach attendance at El Rodadero, Colombia

Abstract

Having accurate data on the number of beach users is vital for planning services and managing overcrowding at high-demand tourist destinations. Obtaining this data has traditionally been a time-consuming task. This study proposes a novel approach to automating the counting process based on computer vision and artificial intelligence. Three HIKVISION PTZ (“Point, Tilt, Zoom”) surveillance cameras were installed on the roofs of the strategic buildings of El Rodadero, Santa Marta, D.T.C.H. Colombia. This observation system provided more than 150 000 observations spanning more than a year (September 2022–October 2023). An algorithm based on the YOLOv5 architecture was trained using a dataset containing over 50 000 labels of people at the beach or in water backgrounds. Applied to surveillance camera images in El Rodadero, Santa Marta, D.T.C.H. Colombia., this algorithm had the ability to identify individuals with high spatiotemporal detail. Beach users were detected over water and on the beach with true positive rates of 81% and 89%, respectively. The counts of beach users made by the model were highly correlated with counts verified by humans with an R2 of 91%. When compared to these expected numbers, the model counted an average of 2–18% fewer people, with a lower error on the beach than on the water. Adjusting for that underestimation the average occupancy at El Rodadero was 125 132 m2 per person, but during peak times, such as long weekends, this number decreased to less than 3 m2 per person. This information is critical for authorities to make informed decisions and for visitors to choose less crowded beaches, avoiding overcrowding. To facilitate access to this information, an automated web system had been created that displayed updated results and allowed downloads for registered users.