摘要
在多目标检测中,当目标发生移动时会导致目标形态变化或目标发生遮挡,影响目标检测精度,因此,针对在自动驾驶领域中行驶环境复杂以及传统目标检测算法依赖大量预设置先验框、泛化能力差、检测精度低等问题,提出了一种改进的CenterNet目标检测算法。选取DLA-34为主干特征提取网络,并引入可自适应确定卷积核且能跨通道交互的轻量级模块ECA-Net,实现CenterNet改进。在kitti数据集上的实验结果显示,改进后的CenterNet相比原网络的AP在car类别上提升了1.39%,在pedestrian类别上提升了11.16%,在cyclist类别上提升了19.45%,与Yolov3网络相比,改进后的CenterNet在不同类别目标检测精度上也有着明显提升。
In multi-target detection,when the target moves,it can cause changes in its shape or occlusion,which affects the accuracy of target detection.Therefore,to address the complex driving environment in the field of autonomous driving and the dependence of traditional object detection algorithms on a large number of pre-set prior boxes,poor generalization ability,and low detection accuracy,an improved CenterNet object detection algorithm is proposed.DLA-34 is selected as the backbone feature extraction network,and a lightweight module ECA-Net that can adaptively determine convolution kernels and interact across channels is introduced to achieve CenterNet improvement.The experimental results on the kitti dataset showed that the improved CenterNet improved the AP of the original network by 1.39%in the car category,11.16%in the pedestrian category,and 19.45%in the cyclist category.Compared with the Yolov3 network,the improved CenterNet also showed significant improvements in object detection accuracy in different categories.
作者
刘凯
宋小军
LIU Kai;SONG Xiaojun(School of Electronic and Information Engineering,Shanghai Electric Power University,Shanghai 200120,China)
出处
《自动化与仪表》
2024年第10期108-112,共5页
Automation & Instrumentation
