基于多尺度特征融合的YOLOv3行人检测算法
YOLOv3 Pedestrian Detection Algorithm Based on Multi-scale Feature Fusion
黎国斌 1王准 1张剑 1扈健玮 1林向会 1谢本亮1
作者信息
- 1. 贵州大学大数据与信息工程学院半导体功率器件可靠性教育部工程研究中心 贵阳 550025
- 折叠
摘要
随着深度学习技术在计算机领域的普及与推广,行人检测技术得到进一步的提升,但在一些场合仍然存在较大的问题,例如行人大小尺度不一、密集行人的检测,在以上两种情况下,行人检测性能剧烈下降,存在较多的漏检、错检的情况.针对行人大小尺度问题,论文提出在YOLOv3算法的特征提取网络中引入多尺度特征融合模块,改变原来多个卷积层堆叠的残差单元,增加特征提取网络深度,提升网络对不同尺度的行人特征提取能力,从而提升行人检测算法的检测精度和鲁棒性.实验表明,在Caltech、On_merge数据集进行训练,改进算法的平均精准率比基准算法分别高出其5.49%,2.26%.
Abstract
With popularization and promotion of deep learning techniques in the field of computer,the pedestrian detection technology has been further improved,but still on several occasions there is a big problem,for example the pedestrian size differ-ence,dense pedestrian detection,in the above two cases,the pedestrian detection performance fell sharply,there exist residual sit-uation and false detection.For pedestrians size problem,YOLOv3 algorithm is introduced in the feature extraction of network multi-scale feature fusion module,changing the original multiple convolution of residual layer stack unit,increasing network depth of feature extraction and improving the network of the different scales of pedestrian feature extraction ability,so as to improve the pe-destrian detection accuracy and robustness of the algorithm.Experimental results show that the average accuracy of the improved al-gorithm is 5.49%and 2.26%higher than that of the benchmark algorithm after training in Caltech and ON_MERGE data sets.
关键词
多尺度特征融合/YOLOv3算法/行人大小尺度/行人检测Key words
multi-scale feature fusion/YOLOv3 algorithm/pedestrian size scale/pedestrian detection引用本文复制引用
基金项目
国家自然科学基金(61562009)
贵州大学引进人才科研项目(贵大人基合字201529号)
半导体功率器件教育部工程研究中心开放基金(ERCMEKFJJ2019-06)
出版年
2024
国家科技期刊平台
NSTL
万方数据