• Laser & Optoelectronics Progress
  • Vol. 57, Issue 10, 101006 (2020)
Qianqian He, Rongfen Zhang, and Yuhong Liu*
Author Affiliations
  • College of Big Data and Information Engineering, Guizhou University, Guiyang, Guizhou 550025, China
  • show less
    DOI: 10.3788/LOP57.101006 Cite this Article Set citation alerts
    Qianqian He, Rongfen Zhang, Yuhong Liu. Human Detection Algorithm Optimization in Machine Vision[J]. Laser & Optoelectronics Progress, 2020, 57(10): 101006 Copy Citation Text show less
    Block diagram of accurate target positioning
    Fig. 1. Block diagram of accurate target positioning
    Algorithm block diagram
    Fig. 2. Algorithm block diagram
    Human detection network model
    Fig. 3. Human detection network model
    Dynamic variation of IOU at different scales
    Fig. 4. Dynamic variation of IOU at different scales
    Loss variation curve
    Fig. 5. Loss variation curve
    Recognition results of darknet. (a) Original border box; (b) body image pre-processed by border frame
    Fig. 6. Recognition results of darknet. (a) Original border box; (b) body image pre-processed by border frame
    Segmentation effect diagrams of different iteration times. (a) Number of iteration is 4; (b) number of iteration is 5; (c) number of iteration is 10
    Fig. 7. Segmentation effect diagrams of different iteration times. (a) Number of iteration is 4; (b) number of iteration is 5; (c) number of iteration is 10
    Comparison of human body effect for image 67-76 in dataset. (a1)-(j1) Proposed algorithm; (a2)-(j2) darknet only
    Fig. 8. Comparison of human body effect for image 67-76 in dataset. (a1)-(j1) Proposed algorithm; (a2)-(j2) darknet only
    IOU contrast graph of three algorithms
    Fig. 9. IOU contrast graph of three algorithms
    Image 91 segmentation effect diagram.(a)Original image segmentation of RGB image; (b) RGB image segmentation of boundary box coordinate processing detected by darknet; (c) after preprocessing the coordinates detected by darknet, the RGB map corresponds to the depth map segmentation
    Fig. 10. Image 91 segmentation effect diagram.(a)Original image segmentation of RGB image; (b) RGB image segmentation of boundary box coordinate processing detected by darknet; (c) after preprocessing the coordinates detected by darknet, the RGB map corresponds to the depth map segmentation
    Image 92 segmentation effect.(a)Original image segmentation of RGB image; (b)RGB image segmentation of boundary box coordinate processing detected by darknet; (c)after preprocessing the coordinates detected by darknet, the RGB map corresponds to the depth map segmentation
    Fig. 11. Image 92 segmentation effect.(a)Original image segmentation of RGB image; (b)RGB image segmentation of boundary box coordinate processing detected by darknet; (c)after preprocessing the coordinates detected by darknet, the RGB map corresponds to the depth map segmentation
    Image 93 segmentation effect.(a)Original image segmentation of RGB image; (b)RGB image segmentation of boundary box coordinate processing detected by darknet; (c)after preprocessing the coordinates detected by darknet, the RGB map corresponds to the depth map segmentation
    Fig. 12. Image 93 segmentation effect.(a)Original image segmentation of RGB image; (b)RGB image segmentation of boundary box coordinate processing detected by darknet; (c)after preprocessing the coordinates detected by darknet, the RGB map corresponds to the depth map segmentation
    ImageIdentification box coordinates /pixelIOU value comparison
    DarknetmodelHOG-SVMalgorithmProposedalgorithmReferencevaluesDarknetHOG+SVMProposedalgorithm
    67(330,185,91,206)(234,58,125,250)(320,175,93,192)(349,97,61,233)0.36950.05320.4209
    68(329,183,83,207)(234,58,125,250)(319,173,93,191)(349,97,61,231)0.39690.05340.4250
    69(326,179,82,215)(234,125,100,200)(316,169,94,194)(344,98,59,228)0.389700.4161
    70(325,178,78,220)(234,125,100,200)(315,168,88,199)(331,99,58,224)0.38940.02430.4208
    71(322,181,79,216)(280,200,64,128)(312,171,88,199)(331,99,58,223)0.37740.08660.4067
    72(319,180,80,217)(218,125,100,200)(309,170,92,199)(326,99,58,225)0.38130.12280.4013
    73(309,185,86,204)(218,125,100,200)(299,175,98,192)(326,99,57,225)0.355800.3700
    74(311,183,79,207)(218,125,100,200)(301,173,91,194)(323,100,57,217)0.365000.3793
    75(308,184,76,203)(218,125,100,200)(298,174,86,191)(323,100,57,216)0.374800.3950
    76(307,188,71,193)(218,125,100,200)(297,178,73,185)(310,101,56,222)0.40960.05740.4584
    Table 1. Three algorithms comparison for detecting human body
    ImageProcessing of imagesNumber of iterationsElapsed time /s
    91Use only the RGB images identified by darknet
    for segmentation
    203.266001
    Preprocessing and segmentation of RGB images
    identified by darknet
    201.894092
    RGB image identified by darknet corresponds to the
    segmentation of depth map
    50.758673
    92Use only the RGB images identified by darknet
    for segmentation
    203.265913
    Preprocessing and segmentation of RGB images
    identified by darknet
    201.904822
    RGB image identified by darknet corresponds to the
    segmentation of depth map
    50.733490
    93Use only the RGB images identified by darknet
    for segmentation
    203.242088
    Preprocessing and segmentation of RGB images
    identified by darknet
    201.867070
    RGB image identified by darknet corresponds to the
    segmentation of depth map
    50.721397
    Average timeUse only the RGB images identified by darknet
    for segmentation
    203.256004
    Preprocessing and segmentation of RGB images
    identified by darknet
    201.888661
    RGB image identified by darknet corresponds to the
    segmentation of depth map
    50.737853
    Table 2. running time of the experiment
    Qianqian He, Rongfen Zhang, Yuhong Liu. Human Detection Algorithm Optimization in Machine Vision[J]. Laser & Optoelectronics Progress, 2020, 57(10): 101006
    Download Citation