摘要
针对自动驾驶情景下行人目标检测过程中对于重叠和遮挡目标存在的漏检问题,提出一种改进多尺度网络YOLOv5的行人目标检测算法.首先,构建同时考虑通道间关系和特征空间位置信息的多重协调注意力模块,增加网络特征表达能力;然后,将原损失函数改进为具有双重惩罚项的切比雪夫距离交并比损失函数,提高检测框的精确度与网络收敛速度;最后,在网络结构方面设计瓶颈状DSP1_X和DSP2_X模块减少梯度混淆.实验结果表明,改进后的多尺度网络收敛能力提高,在面对行车中复杂行人目标检测时具有较高的判别精度和实时检测速度.
Aiming at the missed detection of overlapping and occluded targets in the process of pedestrian target detection in autonomous driving scenarios,a pedestrian target detection algorithm based on improved multi-scale network YOLOv5 is proposed.Firstly,construct a multiple coordinated attention module that considers the relationship between channels and the location information of the feature space at the same time to increase the ability of network feature expression.Then the original loss function is improved to the Chebyshev distance intersection and ratio loss function with double penalty terms to improve the detection frame.The accuracy of the network and the speed of network convergence.Finally,the bottleneck-shaped DSP1_X and DSP2_X modules are designed in the network structure to reduce gradient confusion.The experimental results show that the improved multi-scale network has improved convergence ability,and has higher discrimination accuracy and real-time detection speed when facing complex pedestrian target detection in driving.
作者
罗舜
于娟
LUO Shun;YU Juan(College of Economics and Management,Fuzhou University,Fuzhou,Fujian 350108,China)
出处
《福州大学学报（自然科学版）》
CAS
北大核心
2022年第5期587-594,共8页
Journal of Fuzhou University(Natural Science Edition)
基金
国家自然科学基金资助项目(71771054)。
作者简介
通信作者:于娟(1981-),教授,主要从事数据挖掘与智能信息系统方面研究,yujuan@fzu.edu.cn。
