Intelligent unmanned aerial vehicles surveillance for detecting abnormalities in crowds under occlusion and multi target conditions

Imagery from unmanned aircraft is utilized in this study to present a deep learning framework for the aim of identifying anomalies that has been optimized with the Dragonflies Optimisation Algorithm (DL-DOA) for crowd surveillance in locations with high density. The suggested technique attains a 98.7% F1-score, a 98.5% ROC-AUC, and a low rate of error (MAE 0.03) when assessed using the VisDrone dataset. The suggested approach is proven to be more accurate and resilient than the approaches that are currently in use, as seen by this. The application of sophisticated preprocessing and occluded management in conjunction with YOLO and the Generational adversarial Occlusion Networks (GAON) can potentially result in efficient noise reduction (83–91%) and successful recognition in the presence of partial occlusion. The effectiveness of DL-DOA’s models is enhanced by adaptive hyperparameter adjustment, which maintains a near real-time inference rate of 0.18 s per frame. Despite the fact that there are limitations in extreme weather and extremely dense crowds, this method demonstrates promise for changing, large-scale monitoring situations. Parallel processing will be the focus of future endeavors in order to improve both adaptability and efficiency.