摘要
针对SSD算法对远距离车辆检测中无法考虑局部细节、纹理特征等,进而造成检测精度低的问题,提出了一种基于特征融合的SSD方法并用于远距离车辆检测,将高分辨率的浅层卷积层和语义较强的深层卷积层进行融合,通过残差块设计一个完整的特征融合结构,增加网络的宽度和深度。对KITTI和TT100k数据集进行了检测能力和平均精度均值的对比实验,结果表明,针对远距离车辆,改进后基于特征融合SSD方法的准确率有明显提高。
For single shot detector(SSD)algorithm can not consider local details and texture features in long-distance vehicle detection,which results in low detection accuracy.This paper proposes an improved vehicle detection method.The shallow convolution layer with high resolution and the deep convolution layer with strong semantics are fused.A complete feature fusion structure is designed by residual blocks to increase the width and depth of the network.The comparison experiment of detection ability and average accuracy between KITTI and TT100 kdatasets shows that the improved detection accuracy based on the feature fusion SSD method is significantly improved for long-distance vehicles.
作者
刘鸣瑄
刘惠义
Liu Mingxuan;Liu Huiyi(School of Computer and Information,Hohai University,Nanjing 221100,China)
出处
《国外电子测量技术》
2020年第2期28-32,共5页
Foreign Electronic Measurement Technology