• Acta Optica Sinica
  • Vol. 40, Issue 12, 1215001 (2020)
Yingchun Sun, Shuguo Pan*, Tao Zhao, Wang Gao, and Jiansheng Wei
Author Affiliations
  • School of Instrument Science and Engineering, Southeast University, Nanjing, Jiangsu 210096, China
  • show less
    DOI: 10.3788/AOS202040.1215001 Cite this Article Set citation alerts
    Yingchun Sun, Shuguo Pan, Tao Zhao, Wang Gao, Jiansheng Wei. Traffic Light Detection Based on Optimized YOLOv3 Algorithm[J]. Acta Optica Sinica, 2020, 40(12): 1215001 Copy Citation Text show less
    Relative parameters of Darknet-53
    Fig. 1. Relative parameters of Darknet-53
    Distribution of traffic light label
    Fig. 2. Distribution of traffic light label
    Result of clustering analysis
    Fig. 3. Result of clustering analysis
    Optimized YOLOv3 network structure
    Fig. 4. Optimized YOLOv3 network structure
    Average loss function curve of optimized YOLOv3 algorithm
    Fig. 5. Average loss function curve of optimized YOLOv3 algorithm
    Average IOU curve of optimized YOLOv3 algorithm
    Fig. 6. Average IOU curve of optimized YOLOv3 algorithm
    Detection results of traffic lights by each network. (a)(c) Detection results of YOLOv3 network; (b)(d) detection results of optimized YOLOv3 network
    Fig. 7. Detection results of traffic lights by each network. (a)(c) Detection results of YOLOv3 network; (b)(d) detection results of optimized YOLOv3 network
    AP curve of YOLOv3 network
    Fig. 8. AP curve of YOLOv3 network
    AP curve of optimized YOLOv3 network
    Fig. 9. AP curve of optimized YOLOv3 network
    Comparison of average precision under different scenes and object sizes
    Fig. 10. Comparison of average precision under different scenes and object sizes
    Comparison of detection results. (a) Detection results of YOLOv3 algorithm; (b) detection results of optimized YOLOv3 algorithm
    Fig. 11. Comparison of detection results. (a) Detection results of YOLOv3 algorithm; (b) detection results of optimized YOLOv3 algorithm
    k=4k=5k=6k=7k=8k=9
    (5, 12)(4, 12)(4, 12)(3, 12)(4, 10)(3, 11)
    (5, 21)(5, 20)(6, 18)(5, 18)(5, 16)(5, 13)
    (8, 18)(9, 27)(6, 11)(6, 10)(7, 12)(5, 21)
    (13, 39)(7, 14)(7, 30)(7, 17)(7, 22)(7, 12)
    (14, 47)(12, 21)(8, 33)(8, 33)(8, 18)
    (14,47)(12, 21)(12, 46)(8, 31)
    (15, 49)(13, 22)(12, 45)
    (19, 56)(14, 23)
    (19, 57)
    Table 1. Clustering results under different k values
    AlgorithmTime /msSpeed /(frame·s-1)Precision /%Recall /%AP /%
    YOLOv3372766.5149.2837.67
    Optimized YOLOv3 algorithm333069.7057.5846.78
    Table 2. Comparison of total object detection results
    NetworkBig objectSmall object
    P /%R /%AP /%P /%R /%AP /%
    YOLOv383.9569.3266.5354.5937.9324.00
    Optimized YOLOv386.7175.6872.5456.5844.4829.84
    Table 3. Detection precision, recall, and average precision under different object sizes
    NetworkNight sceneDaytime sceneRain sceneNon-rain scene
    P/%R/%AP/%P/%R/%AP/%P/%R/%AP/%P/%R/%AP/%
    YOLOv365.3649.9837.6770.4544.0436.1570.2851.0041.1863.7147.2934.77
    Optimized YOLOv368.3257.2444.5872.0659.1149.9673.7958.8649.8970.0157.1045.15
    Table 4. Detection precision, recall, and average precision under different scenes
    Yingchun Sun, Shuguo Pan, Tao Zhao, Wang Gao, Jiansheng Wei. Traffic Light Detection Based on Optimized YOLOv3 Algorithm[J]. Acta Optica Sinica, 2020, 40(12): 1215001
    Download Citation