Enhancement of YOLOv8 model for dense crowd scenes: incorporating an improved feature pyramid with attention mechanisms

In the context of complex backgrounds in dense crowd images, small object detection in dense crowds from drone perspectives or surveillance videos faces numerous challenges. These challenges include small target sizes, dense distribution, low resolution, and susceptibility to background noise, making detection significantly difficult. To confront these problems, we implemented a range of enhancements and refinements to the original YOLOv8 model (subsequent articles refer to it as v8 model), proposing a new model named EBE-v8, which stands for ECA-BiFPN-EIoU-YOLOv8. First, the Efficient Channel Attention (ECA) mechanism was integrated into the v8 backbone by us. This operation enhances the model's focus on low-level features, further enhanced the detection capability for small objects. Second, we replaced PANet (the Path Aggregation Network) with a BiFPN (Weighted Bi-directional Feature Pyramid Network) structure to effectively integrate multi-level features, thus increasing the model's capacity to handle targets of different sizes. Finally, we employed a new IoU loss calculation method, EIoU, to better adapt to small object detection. This loss function can dynamically adjust the proportions of different parts of the loss function during various training stages, allowing small targets to better regress to the ground truth boxes and improving small object detection performance. Experimental results demonstrate that the model we enhanced significantly surpasses the initial v8-model in small object detection tasks in dense crowds. The test results on the identical datasets show that our model has achieved an accuracy of 86%, 72.8% recall, 81.6% [email protected], meanwhile [email protected]:0.95 reached 53.5%. Compared to the initial v8-model, these results represent significant improvements, highlighting the efficacy and practicality of our approach.