摘要
目的:针对当前行人目标检测算法中存在的一些产生误检和漏检的情况,提出基于YOLOv7改进的目标检测算法。方法:首先,采用通道注意力机制与YOLOv7中的聚合网络(Efficient Layer Aggregation Networks,ELAN)模块融合,改进成S-ELAN(Squeeze-ELAN)模块,用于强调检测目标的特征重要性;然后,采用空洞空间卷积池化金字塔(Atrous Spatial Pyramid Pooling,ASPP)替换掉原始算法主干网络中的SPPCSPC(Spatial Pyramid Pooling and Cross Stage Partial Connections)模块,用于增强感受野;最后,使用WIoU(Wise Intersection over Union)损失函数对训练模型进行优化。结果:改进后的YOLOv7算法对行人目标识别准确率达到90.26%,相较于普通YOLOv7算法提高1.64%,平均精确度达到61.3%,相较于YOLOv7算法提高2.06%,检测速度相较于原模型提升2.86 f/s。结论:改进后的算法模型,整体泛化能力和性能得到了较大提升,可以较好地在行人检测场景中使用。
Objective:Aiming at the problems of false detection and missing detection in current pedestrian target detection algorithms,an improved target detection algorithm based on YOLOv7 was proposed.Methods:Firstly,the channel attention mechanism was fused with the Efficient Layer Aggregation Networks(ELAN)module in YOLOv7,and improved into S-ELAN(Squeeze-ELAN)module,which was used to emphasize the importance of the characteristics of detection targets.Then,the SPPCSPC(Spatial Pyramid Pooling and Corss Stage Partial Connections)module in the original algorithm backbone network was replaced by Atrous Spatial Pyramid Pooling(ASPP)to enhance the receptive field.Finally,the WIoU(Wise Intersection over Union)loss function was used to optimize the training model.Results:The experimental results showed that the improved YOLOv7 algorithm achieved 90.26%accuracy in pedestrian target recognition,which was 1.64%higher than that of ordinary YOLOv7 algorithm,and the average accuracy reached 61.3%,which was 2.06%higher than that of YOLOv7 algorithm.The detection speed had increased by 2.86 f/s compared to the original model.Conclusion:The improved algorithm model had significantly improved overall generalization ability and performance,and could be well used in pedestrian detection scenarios.
作者
马宇晴
张珂
朱礼龙
谢进
MA Yuqing;ZHANG Ke;ZHU Lilong;XIE Jin(School of Artificial Intelligence and Big Data,Hefei University,Hefei 230601,China)
出处
《安徽科技学院学报》
2024年第5期65-72,共8页
Journal of Anhui Science and Technology University
基金
国家自然科学基金青年项目(12001151)
安徽省自然科学基金(2008085MF202)。
