欢迎访问《图学学报》 分享到：
图学学报

图学学报 ›› 2022, Vol. 43 ›› Issue (2): 288-295.DOI: 10.11996/JG.j.2095-302X.2022020288

• 图像处理与计算机视觉 • 上一篇    下一篇

基于透视降采样和神经网络的地面标志检测

  

  1. 1. 山东大学信息科学与工程学院，山东 青岛 266237；
    2. 山东高速信息集团有限公司，山东 济南 250000
  • 出版日期:2022-04-30 发布日期:2022-05-07
  • 基金资助:

    山东省科技发展计划重点项目(2019GGX101018)；

    山东省自然科学基金项目(ZR2017MF057)

Landmark detection based on perspective down-sampling and neural network

  1. 1. School of Information Science and Engineering, Shandong University, Qingdao Shandong 266237, China;
    2. Shandong Hi-Speed Information Group Co, Ltd, Jinan Shandong 250000, China
  • Online:2022-04-30 Published:2022-05-07
  • Supported by:

    Key Projects of Science and Technology Development Plan of Shandong Province(2019GGX101018); 

    National Natural Science Foundation of Shandong (ZR2017MF057)

PDF (PC)

42

摘要： 在智能驾驶领域，为实时精确检测路面的导向标志，提出一种基于透视降采样和神经网络的地标
检测方法，有效解决传统检测方法实时性较差、复杂场景和远处小目标检测准确率较低的问题。首先，选取图像
感兴趣区域进行透视降采样，降低道路图像近处分辨率，缩小图像尺寸，同时消除透视投影误差。其次对
YOLOv3-tiny 目标检测网络进行改进，采用 k-means++算法对自建数据集的边界框聚类；添加卷积层强化浅层特
征，提升小目标表征能力；改变特征金字塔融合尺度，将预测输出调整为适合地标尺寸的 26×26 和 52×52。最后，
在自建多场景数据集上测试，准确率由 78%提升到 99%，模型大小由 33.8 MB 减小为 8.3 MB。结果表明，基于
透视降采样和神经网络的地标检测方法鲁棒性强，对小目标检测精度更高，易于在低端嵌入式设备上部署。

关键词: 透视降采样, YOLOv3-tiny, 地标检测, 数据集, k-means++

Abstract: In the field of intelligent driving, a neural network-based and perspective down-sampling-based landmark
detection method was proposed to accurately detect the road guide signs in real time. This proposed method can
effectively solve the problems of poor real-time performance of traditional detection methods and low detection
accuracy for complex scenes and remote small targets. Firstly, the region of interest for the image was selected for
perspective down-sampling to reduce the near resolution of the road image, reduce the image size, and eliminate the
perspective projection error. Secondly, the YOLOv3-tiny target detection network was enhanced. The boundary frame
clustering of self-built data set was implemented by k-means++. The convolution layer was added to strengthen the
shallow features and enhance the small target representation ability. By changing the fusion scale of feature pyramid,
the prediction output was adjusted to 26×26 and 52×52. Finally, the accuracy rate was elevated from 78% to 99% on
the self-built multi-scene data set, and the model size was reduced from 33.8 MB to 8.3 MB. The results show that a neural network-based and perspective down-sampling-based landmark detection method displays strong robustness,
higher detection accuracy for small targets, and is readily deployable on low-end embedded devices.

Key words: perspective down-sampling, YOLOv3-tiny, landmark detection, data set, k-means++

中图分类号: