Research on Statistical Counting Model of Overlapping Scene Based on YOLO

Aiming at the requirement of people counting in complex surveillance scenarios, this study proposes a two-stage solution that fuses the improved YOLO detector with cross-modal re-recognition. First, in the single-camera detection stage, YOLO-11x's detection capability for occluded targets is enhanced by improved mosaic data augmentation and adaptive color perturbation strategies. Second, in the multi-camera de-recognition stage, a semantic-aware matching model is designed to be combined with a dynamic threshold decision-making mechanism based on the minimization of KL scatter. Experiments show that the optimized detection model reaches 97.3% on the self-built dataset. This study provides a highly robust people counting scheme for real-world surveillance scenarios of cross-area people counting in large scenes.