作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• •    

面向拥挤行人检测的改进YOLOv7算法

  • 发布日期:2023-10-12

Improved YOLOv7 algorithm for crowded pedestrian detection

  • Published:2023-10-12

摘要: 拥挤行人检测一直是行人检测领域的研究热点。针对拥挤行人检测场景下检测算法易产生漏检与误检的问题,提出了一种改进的YOLOv7目标检测算法。针对拥挤行人检测场景中被遮挡行人目标特征缺失的问题,在骨干网络中融入Bi-Former视觉变换器模块以及改进的RC-ELAN模块,通过引入自注意力机制与注意力模块使骨干网络更多聚焦于被遮挡行人重要特征,有效缓解特征缺失对检测造成的影响。针对拥挤行人检测场景中小目标行人易被漏检的问题,采用融入BIFPN思想的改进颈部网络,通过引入转置卷积以及改进的Rep-ELAN-W模块使模型可以高效利用中低维特征图中的小目标特征信息,有效提升模型的小目标行人检测性能。针对原损失函数训练效率较低的问题,引入Efficient-CIoU损失函数,使模型可以进一步收敛至更高的精度。最后在含有大量小目标遮挡行人的WiderPerson拥挤行人检测数据集上进行实验表明,改进后的YOLOv7算法在拥挤行人检测场景下领先YOLOv7算法0.025AP50以及0.028AP50:95精度,领先YOLOv5算法0.099AP50以及0.071AP50:95精度,领先YOLOX算法0.123AP50以及0.107AP50:95精度。所提算法可以较好地应用于拥挤行人检测场景。

Abstract: Crowded pedestrian detection has been a hot research topic in the field of pedestrian detection. An improved YOLOv7 target detection algorithm is proposed for crowded pedestrian detection scenarios where the detection algorithm is prone to miss and false detection. To address the problem of missing target features of obscured pedestrians in crowded pedestrian detection scenarios, the Bi-Former visual transformer module and the improved RC-ELAN module are incorporated into the backbone network, and the Self-Attention mechanism and attention module are introduced to make the backbone network focus more on important features of obscured pedestrians and effectively mitigate the impact of missing features on detection. To address the problem that small target pedestrians are easily missed in crowded pedestrian detection scenarios, an improved neck network incorporating the idea of BIFPN is used. By introducing transposed convolution and an improved Rep-ELAN-W module, the model can efficiently utilize the small target feature information in the low and medium dimensional feature maps to effectively improve the small target pedestrian detection performance of the model. To address the problem of low training efficiency of the original loss function, the Efficient-CIoU loss function is introduced so that the model can be further converged to higher accuracy. Finally, experiments on the WiderPerson crowded pedestrian detection dataset with a large number of small targets obscuring pedestrians show that the improved YOLOv7 algorithm leads the YOLOv7 algorithm by 0.025 AP50 and 0.028 AP50:95 accuracy in crowded pedestrian detection scenarios, and leads the YOLOv5 algorithm by 0.099 AP50 and 0.071 AP50:95 accuracy, ahead of YOLOX algorithm 0.123 AP50 and 0.107 AP50:95 accuracy.