摘要
针对无人驾驶中交通标志尺寸小、占据像素点少、特征不明显而带来的检测困难且无法实时识别的问题,提出一种基于Inception⁃Yolov3(Inc⁃Yolov3)的目标检测算法。首先从Yolov3的网络结构出发,在其浅层网络处嵌入Inception⁃redefined module结构,该结构采用多分支的分布方式来增加网络宽度,多尺度卷积核获得不同尺度的感受野信息以及多个1×1卷积的使用大大减少了网络参数;然后将此结构连接到Yolov3的深层网络中,以此融合深层特征与浅层特征,丰富了用于检测小目标的浅层网络的语义信息,进而增强了浅层网络的特征提取能力;其次通过计算物体坐标的宽高比例来消除数据集中无效数据的方法对Yolov3中原始K⁃means聚类算法进行改进,以获得更加匹配数据集的锚框大小与比例;最后在训练时使用GIOU代替IOU作为坐标损失函数,GIOU具有更加精细的目标尺度并能反映预测框与真实框的各种重叠情况。实验结果表明:与其他经典的检测算法相比,Inc⁃Yolov3算法在满足实时性的条件下,识别精度更高;在TT100K交通标志数据集上获得了86.1%的准确率。
The small size,few pixel occupation and inconspicuous features of the traffic signs in unmanned driving lead to difficult detection and failure of real⁃time recognition in unmanned driving,so a target detection algorithm based on Inception⁃Yolov3(Inc⁃Yolov3)is proposed.On the basis of the network structure of Yolov3,the Inception⁃redefined module structure is embedded in its shallow network.In this structure,a multi⁃branch distribution method is adopted to increase the network width,the multi⁃scale convolution kernel is used to obtain different scales of receptive field information and the use of 1×1 convolutions reduces the network parameters greatly.And then,the structure is connected to the deep network of Yolov3 to fuse deep features and shallow features,which enriches the semantic information of the shallow network used to detect small targets and further enhances the feature extraction capabilities of the shallow network.Then the original K⁃means clustering algorithm in Yolov3 is improved by calculating the width and height ratio of the object coordinates to eliminate the invalid data in the dataset,so as to obtain the anchor frame size and proportion that more closely match the dataset.At last,GIOU(generalized intersection over union)is used as the coordinate loss function instead of IOU(intersection over union)in training.GIOU has a finer object scale and can reflect various overlaps between the predicted frame and the real frame.The experimental results show that the Inc⁃Yolov3 algorithm has higher recognition accuracy under real⁃time conditions in comparison with the other classic detection algorithms.An accuracy rate of 86.1%is obtained on the TT100K traffic sign dataset.
作者
马新舒
唐欣
陈艳
姚荣彬
李晓欢
MA Xinshu;TANG Xin;CHEN Yan;YAO Rongbin;LI Xiaohuan(School of Information and Communication,Guilin University of Electronic Technology,Guilin 541004,China;Guilin Institute of Information Technology,Guilin 541004,China;School of Computer Science and Engineering,Guilin University of Aerospace Technology,Guilin 541004,China)
出处
《现代电子技术》
2022年第13期179-186,共8页
Modern Electronics Technique
基金
国家自然科学基金项目(61762030)
广西自然科学基金项目(2019GXNSFFA245007,2018GXNSFDA281013)
广西科技计划项目(AA18242021,AB19110050,AA19110044,ZY19183005,AB20238033)
桂林市科技计划项目(20190214⁃3)
广西高校中青年教师基础能力提升项目(2018KY0651)。