基于改进YOLOv7的道路目标检测方法

Road target detection method based on improved YOLOv7

针对现有方法在道路场景中实施目标检测时存在对小目标检测精度低、泛化性能不佳等问题,提出一种基于YOLOv7的改进方法。在特征融合结构中,引入通道注意力机制来抑制更多负样本参与特征学习,同时在融合层末端输出四个尺寸的特征图,以强化对小尺寸目标的检测精度;使用改进K均值聚类(K-means++)算法生成与真实目标宽高更贴合的锚点框,让模型在训练前期快速收敛;最后在检测输出端使用软性非极大值抑制算法,精准检测重叠度较高的目标。以开源中国交通标志数据集(CCTSDB)与腾讯-清华100K(TT100K)数据集混合构建训练与测试数据集,实验结果表明,与原始YOLOv7相比,改进后模型在m_(AP)@0.5、m_(AP)@0.5:0.95指标上分别提升7.9%与5.6%,同时检测速度仅有少量下降,但仍能完成实时检测,同时在不同场景下保持性能稳定,充分证明了本文所提方法能够在复杂道路场景下开展多类目标的快速精准检测。

To solve the problems of low detection accuracy and poor generalization performance of the existing methods for target detection in road scenarios,an improved method based on YOLOv7 was proposed.In the feature fusion structure,the channel attention mechanism was introduced to suppress more negative samples from participating in feature learning and output feature maps of four sizes at the end of the fusion layer,so as to enhance the detection accuracy of small-size targets.The improved K-means clustering algorithm++(K-means++)was used to generate the anchor point frame with a closer fit of the width and height of the real target,so as to allow the model to quickly converge in the early stage of training.Finally,a soft non-maximum suppression algorithm was used at the detection output end to accurately detect targets with a high degree of overlap.The training and testing data sets were constructed by mixing the open source China traffic sign data set(CCTSDB)and the Tencent-Tsinghua 100K(TT100K)data set.The experimental results show that compared with the original YOLOv7,the improved model improves the mAP@0.5 and_m(AP)@0.5∶0.95 indicators by 7.9%and 5.6%,respectively,and the detection speed only drops slightly,but it can still realize real-time detection.At the same time,the stable performance in different scenarios fully proves that the method proposed in this paper can quickly and accurately detect multiple types of targets in complex road scenarios.