Improving the accuracy of drowning detection based on improved YOLOv5

Drowning stands as a primary cause of unintentional deaths globally. This paper presents an improved YOLOv5 algorithm tailored for drowning detection, aiming to effectively mitigate drowning incidents. The improved YOLOv5 incorporates the Ghost-CBAM-C3 (GCC) module, which comprises Ghost-bottleneck modules and the CBAM module, and the learning rate decay of Cosine Annealing. To gauge the algorithm's efficacy, four self-made data sets were curated utilising a DJI mini3pro drone over both swimming pools and natural water bodies. Experimental findings underscore the heightened performance of the improved YOLOv5 over the original YOLOv5s. This enhancement manifests in a precision boost from 92.8 to 97.1%, and the values for mean average precision ([email protected]), weights and the Frames-Per-Second (FPS) are 93.2, 14.1 and 23.70, respectively, affirming its applicability in real-time scenarios. Furthermore, results indicate superior performance of the swimming pool data set compared to those from natural water bodies.