基于ASF-YOLOv8的交通场景多目标检测算法

Multi-Object Detection Algorithm for Traffic Scenarios Based on ASF-YOLOv8

摘要: 针对复杂交通环境下的多目标检测问题，本文提出了一种改进YOLOv8的目标检测算法ASF-YOLOv8。首先，在YOLOv8的基础架构上，加入一种注意尺度序列融合机制(Attentional Scale Sequence Fusion, ASF)，该机制能够对不同尺度的特征图进行融合，从而获得更好的图像特征，提取出更丰富、更准确的特征信息。然后，对损失函数进行改进，引入Inner-IoU，通过辅助边框计算IoU损失，进一步提高算法的检测精度。实验结果表明，在VisDrone数据集上，本文所提算法比YOLOv8算法的平均精度mAP50提升了1.4%，该算法在复杂交通环境下具有更高的检测精度。

Abstract: For multi-object detection problems in complex traffic environments, this paper proposes an improved YOLOv8 object detection algorithm named as ASF-YOLOv8. Firstly, on the infrastructure of YOLOv8, an Attentional Scale Sequence Fusion, ASF is added, which can fuse feature maps at different scales, so as to obtain better image features and extract richer and more accurate feature information. Then, the loss function is improved by introducing Inner-IoU, which can further improve the detection accuracy of the algorithm by calculating the loss through auxiliary frame. The experimental results show that the detection accuracy mAP50 is improved by 1.4% on VisDrone dataset, so the proposed algorithm has more sufficient detection accuracy in complex traffic environment.