An Optimized YOLOv8 Approach for Pedestrian Detection in High-Density Crowds

To address the issues of low accuracy, high miss rates, excessive computational costs, and large parameter sizes in existing crowd density estimation algorithms, this paper proposes YOLOv8-EM. First, we embed the ECA module (Efficient Channel Attention) at the end of the backbone network. By avoiding dimensionality reduction and employing 1D convolution for local cross-channel interaction, it adaptively determines kernel sizes to capture channel-wise dependencies, enhancing the model’s perception and understanding of targets in high-density crowd scenarios. Second, we design an MSConv (Multi-Scale Convolution) feature extraction module and replace the original C2f structure in the backbone with an improved C2f-MS. This modification reduces parameters and computational overhead while maintaining efficiency, further optimizing feature representation and enhancing feature extraction and fusion capabilities. Finally, we replace the original CIoU loss with Efficient IoU (EIOU), improving detection accuracy for small targets and boosting performance in complex scenes. Experiments on the Wider-person dataset demonstrate that YOLOv8-EM achieves an accuracy of 79.7% and an mAP50 of 71%, outperforming the baseline YOLOv8 by 1.7%.