计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (15): 97-106.DOI: 10.3778/j.issn.1002-8331.2207-0313

• 目标检测专题 • 上一篇    下一篇

基于改进YOLOv4-Tiny轻量化校内行人目标检测算法

孙好,董兴法,王军,陈致远   

  1. 1.苏州科技大学 电子与信息工程学院,江苏 苏州 215009
    2.中国科学院 长春光学精密机械与物理研究所,长春 130033
    3.中国白城兵器试验中心,吉林 白城 137001
  • 出版日期:2023-08-01 发布日期:2023-08-01

Lightweight Intra-School Pedestrian Detection Algorithm Based on Improved YOLOv4-Tiny

SUN Hao, DONG Xingfa, WANG Jun, CHEN Zhiyuan   

  1. 1.School of Electronic and Information Engineering, Suzhou University of Science and Technology, Suzhou, Jiangsu  215009, China
    2.Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Sciences, Changchun 130033, China
    3.Center of Arms Experiment of Baicheng, Baicheng, Jilin 137001, China
  • Online:2023-08-01 Published:2023-08-01

摘要: 深度学习常用于行人检测,为了在嵌入式设备上应用复杂的传统卷积神经网络,网络的轻量化是必然趋势,但难以兼顾速度和精度。为解决这个问题,设计了一种基于改进YOLOv4-Tiny的轻量化校内行人目标检测算法。提出了一种多尺度空洞卷积模块的改进Ghost卷积特征提取模块,同时普通卷积用深度可分离卷积代替,降低了模型复杂度,增加特征提取的多样性;构建了一种空洞深度可分离卷积的改进空间金字塔池化结构,增强上下文特征的融合,提高检测精度,减少网络参数;再引入Soft-NMS取代传统非极大值抑制,降低漏检率。实验结果表明,该算法在多个数据集和硬件平台上,其具有精度高、速度快、模型参数少和体积少等特点,可以应用于嵌入式设备。

关键词: 校内行人, 深度学习, YOLOv4-Tiny, Ghost卷积, 非极大值抑制

Abstract: Deep learning is often used for pedestrian detection. In order to apply complex traditional convolutional neural networks on embedded devices, the lightweight of the network is an inevitable trend, but it is difficult to balance speed and accuracy. To solve this problem, this paper designs a lightweight intra-school pedestrian target detection algorithm based on improved YOLOv4-Tiny. Firstly, an improved Ghost convolution feature extraction module is proposed, which is a multi-scale hole convolution module. At the same time, ordinary convolution is replaced by depthwise separable convolution, which reduces the complexity of the model and increases the diversity of feature extraction. Secondly, an improved spatial pyramid pooling structure with depthwise separable convolution of holes enhances the fusion of contextual features, improves detection accuracy, and reduces network parameters. Finally, Soft-NMS is introduced to replace traditional non-maximum suppression to reduce the missed detection rate. Experimental results show that the algorithm has the characteristics of high accuracy, fast speed, few model parameters and small size on multiple data sets and hardware platforms, and can be applied to embedded devices.

Key words: intra-school pedestrians, deep learning;YOLOv4-Tiny, Ghost convolution, non-maximum suppression