• Laser & Optoelectronics Progress
  • Vol. 59, Issue 18, 1815004 (2022)
Zhaoxin Li, Shuhua Lu*, Lingqiang Lan, and Qiyuan Liu
Author Affiliations
  • College of Information and Cyber Security, People’s Public Security University of China, Beijing 102600, China
  • show less
    DOI: 10.3788/LOP202259.1815004 Cite this Article Set citation alerts
    Zhaoxin Li, Shuhua Lu, Lingqiang Lan, Qiyuan Liu. Convolutional Neural Network Method for Crowd Counting Improved using Involution Operator[J]. Laser & Optoelectronics Progress, 2022, 59(18): 1815004 Copy Citation Text show less
    Crowd counting network structure
    Fig. 1. Crowd counting network structure
    Generation and working principle of Involution kernel
    Fig. 2. Generation and working principle of Involution kernel
    Examples of density map generation
    Fig. 3. Examples of density map generation
    Example of the ablation experiment
    Fig. 4. Example of the ablation experiment
    Datasetα
    SHHA1000
    SHHB100
    UCF-QNRF1000
    UCF_CC_50100
    Table 1. Value of coefficient α in different datasets
    MethodSHHASHHBUCF-QNRFUCF_CC_50
    MAEMSEMAEMSEMAEMSEMAEMSE
    MCNN24110.2173.226.441.3243.5364.7467.0498.5
    Switching CNN2690.4135.021.633.4228.0445.0318.1439.2
    CMTL31101.3152.420.031.1252514322.8397.9
    CSRNet1668.2115.010.616.0120.3208.5266.1397.5
    SANet3267.0104.58.413.6----
    ACSPNet1885.2137.115.423.1----
    PCCNet1073.5124.011.019.0149.0247.0240.0315.5
    AMCNN276.1110.715.327.4----
    LSC-CNN3366.4117.08.112.7120.5218.2255.6302.7
    TEDNet3064.2109.18.212.8113.0188.0249.4354.5
    SCLNet2067.9102.99.114.1109.6182.5258.9326.2
    DSPNet1368.2107.88.914.0107.5182.7243.3307.6
    MSCANet3466.5109.4--104.1183.8242.8329.8
    Method in Ref.[1561.9100.57.411.7104.8182.3212.3289.6
    AMS-Net1163.8108.57.311.886.5167.2236.5319.2
    Proposed method61.1101.37.011.3102.5181.7202.0288.7
    Table 2. Comparison results of the different methods
    ComponentMAEMSETime /msParameters /106
    VGG(dilation rate is 2)(baseline)69.1106.2166.5411.54
    VGG+1INV(dilation rate is 2)66.9105.2145.3811.81
    VGG+2INV(dilation rate is 2)64.9101.4133.8512.08
    VGG+3INV(dilation rate is 1)67.0108.5132.2012.35
    VGG+3INV(dilation rate is 2)63.6103.6127.5312.35
    VGG+3INV(dilation rate is 2)+residual connection62.9106.5133.6312.35
    VGG+3INV(dilation rate is 2)+residual connection+Loss61.1101.3136.4312.35
    Table 3. Ablation study on the SHHA dataset
    Zhaoxin Li, Shuhua Lu, Lingqiang Lan, Qiyuan Liu. Convolutional Neural Network Method for Crowd Counting Improved using Involution Operator[J]. Laser & Optoelectronics Progress, 2022, 59(18): 1815004
    Download Citation