• Optics and Precision Engineering
  • Vol. 30, Issue 19, 2390 (2022)
Shuai HAO, Tian HE, Xu MA*, Lei YANG, and Siya SUN
Author Affiliations
  • College of Electrical and Control Engineering, Xi'an University of Science and Technology, Xi'an710054, China
  • show less
    DOI: 10.37188/OPE.20223019.2390 Cite this Article
    Shuai HAO, Tian HE, Xu MA, Lei YANG, Siya SUN. Cross-scale infrared pedestrian detection based on dynamic feature optimization mechanism[J]. Optics and Precision Engineering, 2022, 30(19): 2390 Copy Citation Text show less
    Structure diagram of DFOM-CSNet network
    Fig. 1. Structure diagram of DFOM-CSNet network
    Structure diagram of LPM module
    Fig. 2. Structure diagram of LPM module
    Comparison of images before-and-after infrared feature dynamic optimization
    Fig. 3. Comparison of images before-and-after infrared feature dynamic optimization
    Comparison of feature pyramid structures
    Fig. 4. Comparison of feature pyramid structures
    Structure of cross-scale feature fusion module
    Fig. 5. Structure of cross-scale feature fusion module
    Results of ablation experiment
    Fig. 6. Results of ablation experiment
    Comparison of detection results
    Fig. 7. Comparison of detection results

    算法1动态特征优化机制

    Input: 红外图像Iir,最大优化迭代次数Max_iteration,寻优参数α,搜索种群XAttackerXChaserXBarrierXDriver.

    1. LPM模块

     构造四叉树-贝塞尔插值算子重构初始红外背景图像;

     引入引导滤波平滑噪音并得到红外亮度特征图像ILir和红外背景图像IBir

    2. EG-Chimp优化模型

     构建动态特征优化图像:IOir=α×ILir+IBir

     设计目标函数对参数α寻优:F=min{LSF+λLCON}

    Whilet<Max_iteration

      For each chimp

       计算各人猿种群的位置向量;

       更新f,m,c,a,D

      End For

       For each search chimp

        更新目前搜索种群的位置向量;

       End For

        更新XAttackerXChaserXBarrierXDriver

        t=t+1

    End While

    Output: 动态特征优化图像IOir

    Table 1. [in Chinese]
    ConfigurationVersion parameter
    Operating systemMicrosoft Windows 10
    GPUNVIDIA GeForce GTX 1660 Ti
    CPUIntel Core i5-10400F@2.90 GHz×6 CPUs
    CUDA11.1
    Deep learning frameworkPytorch
    Table 1. Software and hardware platform configuration
    ImageEntropyBrennerDCTVariance
    IR5.654.72×1064.69×1031.06×108
    Optimized IR6.042.84×1074.87×1031.92×108
    Table 2. Average values of evaluation indexes for 1 000 images
    ComponentMethod
    GIOU
    CIOU
    DFOM
    YOLOv5
    CSFF-BiFPN
    mAP@.5/%88.388.889.489.190.289.889.590.7
    Table 3. Improved module validation
    MethodBackbonemAP@.5/%mAP@.5:.95/%Time/s
    Faster-RCNNResNet5072.90.163
    SSDVGG75.10.087
    RetinaNetResNet5074.836.30.140
    Sparse R-CNNResNet5084.342.90.148
    VarifocalNetResNet5086.245.30.172
    TOODResNet5088.946.10.156
    YOLOv4-CLAHECSPDarkNet87.945.90.042
    I-YOLODarkNet88.945.80.024
    TC-DetDarkNet89.246.20.149
    CSNet(Ours)Focus+ CSPDarkNet89.546.20.011
    DFOM-CSNet(Ours)Focus+ CSPDarkNet90.746.80.013
    Table 4. Comparison results of different detection algorithms
    Shuai HAO, Tian HE, Xu MA, Lei YANG, Siya SUN. Cross-scale infrared pedestrian detection based on dynamic feature optimization mechanism[J]. Optics and Precision Engineering, 2022, 30(19): 2390
    Download Citation