• Laser & Optoelectronics Progress
  • Vol. 59, Issue 12, 1215002 (2022)
Yong Xuan1, Chao Han1、*, and Wenhan Sha2
Author Affiliations
  • 1Key Laboratory of Advanced Perception and Intelligent Control of High-End Equipment, Ministry of Education, Anhui Polytechnic University, Wuhu 241000, Anhui , China
  • 2Chery New Energy Automobile Co., Ltd., Wuhu 241000, Anhui , China
  • show less
    DOI: 10.3788/LOP202259.1215002 Cite this Article Set citation alerts
    Yong Xuan, Chao Han, Wenhan Sha. Improved Tiny YOLOv4 Algorithm and Its Application in Pedestrian Detection[J]. Laser & Optoelectronics Progress, 2022, 59(12): 1215002 Copy Citation Text show less
    Network structure of Tiny YOLOv4
    Fig. 1. Network structure of Tiny YOLOv4
    Improved network structure
    Fig. 2. Improved network structure
    Network structures of traditional convolution and depthwise separable convolution. (a) Network structure of traditional convolution; (b) network structure of depthwise separable convolution
    Fig. 3. Network structures of traditional convolution and depthwise separable convolution. (a) Network structure of traditional convolution; (b) network structure of depthwise separable convolution
    Channel attention structure
    Fig. 4. Channel attention structure
    Structure diagram of spatial attention module
    Fig. 5. Structure diagram of spatial attention module
    Feature enhancement module
    Fig. 6. Feature enhancement module
    Loss values of different network structures. (a) Loss values of YOLOv4; (b) loss values of Tiny YOLOv4; (c) loss values of improved network
    Fig. 7. Loss values of different network structures. (a) Loss values of YOLOv4; (b) loss values of Tiny YOLOv4; (c) loss values of improved network
    Experimental results
    Fig. 8. Experimental results
    Detection algorithmNumber of training parameters
    YOLOv464040001
    Tiny YOLOv45939804
    Proposed Tiny YOLOv44369606
    Table 1. Comparison of training parameters of different network models
    Detection algorithmModel size /MB
    YOLOv4246
    Tiny YOLOv422.7
    Proposed Tiny YOLOv415.9
    Table 2. Size comparison of different network models
    DatasetAlgorithmAP /%FPS /(frame·s-1
    INRIAYOLOv486.220.1
    Tiny YOLOv468.6731.6
    Proposed Tiny YOLOv476.3230.4
    COCOYOLOv490.7620.8
    Tiny YOLOv469.7632.4
    Proposed Tiny YOLOv478.230.2
    VOCYOLOv490.419.8
    Tiny YOLOv471.2731.2
    Proposed Tiny YOLOv478.830.6
    Mixed dataYOLOv492.519.5
    Tiny YOLOv473.4232.2
    Proposed Tiny YOLOv480.5231.4
    Table 3. Test results of different algorithms on different datasets
    AlgorithmAP /%Recall /%FPS /(frame·s-1
    Faster R-CNN72.3778.44.7
    SSD78.2079.122.1
    YOLOv385.3477.617.3
    Tiny YOLOv367.8073.425.6
    YOLOv492.5081.519.5
    Tiny YOLOv473.4275.732.2
    Proposed Tiny YOLOv480.5282.331.4
    Table 4. Comparison of results of detection algorithm
    Tiny YOLOv4 baselineDSCAttention mechanismScale enhancementAP /%
    73.42
    74.32
    77.27
    80.52
    Table 5. Ablation experiments on mixed datasets
    Yong Xuan, Chao Han, Wenhan Sha. Improved Tiny YOLOv4 Algorithm and Its Application in Pedestrian Detection[J]. Laser & Optoelectronics Progress, 2022, 59(12): 1215002
    Download Citation