Crowd detection from the perspective of UAV Based on YOLOv8n

In traffic scenes from the perspective of drones, pedestrian targets often exhibit small, irregular, and low resolution characteristics, and are often affected by complex backgrounds and dense occlusion interference, resulting in high rates of missed and false detections in traditional object detection algorithms. To address this issue, we propose an improved model based on YOLOv8n, ODA-YOLOv8n,which significantly improves the detection robustness in drone scenarios through dynamic feature enhancement and multi-scale context modeling. The core improvement is as follows: ODConv (Omni Dimensional Dynamic Convolution) is used instead of static convolution, and the spatial, channel, and input dimension weights of the convolution kernel are dynamically adjusted to enhance the model's feature extraction ability for small targets and occluded areas. Experiments have shown that the attention mechanism of ODConv can effectively suppress background noise and reduce false alarm rates; Then, the PSA extreme attention mechanism was inserted into the NECK network structure to directly fit fine-grained regression distributions through nonlinear combinations, reducing information loss caused by dimensionality reduction. Experiments on the VisDrone2019 dataset showed that ODA-YOLOv8 achieved 89.3% accuracy, 77.5% recall, and 79.5% accuracy mAP@50 Compared to the baseline YOLOv8n, they increased by 6.3%, 4.12%, and 1.93%, respectively.