摘要
针对目前主流的目标检测算法存在模型参数过大、不能很好地移植到移动设备端之中应用于辅助驾驶这一问题,本文提出了一种改进YOLOv5s的目标检测算法。首先,将YOLOv5s算法的主干网络CSPDarknet替换为轻量化网络模型MobileNet-V3,解决了网络模型较大、参数较多的问题,减少了网络的深度并提升了数据推断速度;其次,对特征提取网络采用加权双向特征金字塔结构Bi-FPN加强特征提取,融合多尺度特征进而扩大感受野;最后,对损失函数进行优化,使用CIoU为边界框回归损失函数,改善模型原始GIoU收敛速度较慢问题,使预测框更加符合于真实框,同时降低网络训练难度。实验结果表明,改进后算法在KITTI数据集上的mAP相较于YOLOv5s、SSD、YOLOv3、YOLOv4_tiny分别提升了4.4、15.7、12.4、19.6,模型大小相较于YOLOv5s与轻量级网络YOLOv4_tiny分别减少了32.4 MB和21 MB,同时检测速度分别提升了17.6%与43%。本文改进后的算法满足模型小、精确度高的要求,为辅助驾驶中道路目标检测提升检测速度与精度提供了一种解决方案。
Aiming at the problem that the model parameters of the current mainstream target detection algorithms are too large and cannot be transplanted to mobile devices and applied to assisted driving,this paper proposes an improved YOLOv5s target detection algorithm.Firstly,CSPDarknet,the backbone network of YOLOv5s algorithm,is replaced by MobileNet-V3,a lightweight network model,which solves the problem of large network model and many parameters,reduces the network depth and improves the data inference speed.Secondly,a weighted bidirectional feature pyramid structure Bi-FPN is used to enhance feature extraction,and multi-scale features are integrated to expand the receptive field.Finally,the loss function is optimized and CIoU is used as the boundary box regression loss function to improve the slow convergence speed of the original GIoU model,so that the prediction box is more consistent with the real box,and at the same time reduce the difficulty of network training.Experimental results show that compared with YOLOv5s,SSD,YOLOv3 and YOLOv4_tiny,the mAP of the improved algorithm on KITTI dataset is improved by 4.4,15.7,12.4 and 19.6,respectively.Compared with YOLOv5s and lightweight network YOLOv4_tiny,the model size is reduced by 32.4 MB and 21 MB respectively,and the detection speed is improved by 17.6%and 43%respectively.The improved algorithm meets the requirements of small model and high accuracy,and provides a solution for improving detection speed and accuracy of road target detection in assisted driving.
作者
周晴
谭功全
尹宋麟
李易念
魏丹芹
ZHOU Qing;TAN Gong-quan;YIN Song-lin;LI Yi-nian;WEI Dan-qin(School of Automation and Information Engineering,Sichuan University of Science&Engineering,Zigong 643000,China;Artificial Intelligence Key Laboratory of Sichuan Province,Yibin 644000,China)
出处
《液晶与显示》
CAS
CSCD
北大核心
2023年第5期680-690,共11页
Chinese Journal of Liquid Crystals and Displays
基金
四川省科技厅项目(No.2020JDJQ0075)
人工智能四川省重点实验室科研项目(No.2019RYJ08)。
