• Acta Photonica Sinica
  • Vol. 51, Issue 6, 0610006 (2022)
Shuai HAO1, Shan GAO1, Xu MA1、*, Beiyi AN1, Tian HE1, Hu WEN2, and Feng WANG3
Author Affiliations
  • 1College of Electrical and Control Engineering,Xi'an University of Science and Technology,Xi'an 710054,China
  • 2College of Safety Science and Engineering,Xi'an University of Science and Technology,Xi'an 710054,China
  • 3College of Physics and Electrical Engineering,Weinan Normal University,Weinan,Shaanxi 714000,China
  • show less
    DOI: 10.3788/gzxb20225106.0610006 Cite this Article
    Shuai HAO, Shan GAO, Xu MA, Beiyi AN, Tian HE, Hu WEN, Feng WANG. Infrared Pedestrian Detection Based on Cross-scale Feature Aggregation and Hierarchical Attention Mapping[J]. Acta Photonica Sinica, 2022, 51(6): 0610006 Copy Citation Text show less
    Flow chart of detection algorithm
    Fig. 1. Flow chart of detection algorithm
    YOLOv4 object detection algorithm
    Fig. 2. YOLOv4 object detection algorithm
    Focus slice sampling principle
    Fig. 3. Focus slice sampling principle
    Focus+CBM network structure
    Fig. 4. Focus+CBM network structure
    Spatial pyramid pooling structure
    Fig. 5. Spatial pyramid pooling structure
    Multi-scale feature aggregation module based on spatial pyramid pooling
    Fig. 6. Multi-scale feature aggregation module based on spatial pyramid pooling
    Hierarchical attention mapping module based on CBAM
    Fig. 7. Hierarchical attention mapping module based on CBAM
    Comparison of saliency distribution
    Fig. 8. Comparison of saliency distribution
    Comparison of scale distribution in test results
    Fig. 9. Comparison of scale distribution in test results
    Comparison of visualization detection results
    Fig. 10. Comparison of visualization detection results
    P-R curve comparison of different algorithms
    Fig. 11. P-R curve comparison of different algorithms
    Performance curve of the CFAHAM algorithm
    Fig. 12. Performance curve of the CFAHAM algorithm
    Detection experiment under noise condition
    Fig. 13. Detection experiment under noise condition
    Detection results under occlusion conditions
    Fig. 14. Detection results under occlusion conditions
    ConfigurationVersion parameters
    GPUNvidia GeForce GTX 1080Ti
    CPUIntel(R)Core(TM)i7-8700K 3.70GHz @2.90GHz×6 CPUs
    Operating systemMicrosoft Windows 10
    Deep learning frameworkPytorch 1.2.0 CUDA 10.0
    Table 1. Experimental environment configuration parameters
    K-meansSPPCBAMFocusAP/%Precision/%Recall/%F1
    87.4691.7977.090.84
    91.4887.4385.870.87
    94.7792.7287.880.90
    94.2189.6991.880.91
    95.3794.2592.990.94
    Table 2. Comparison of detection results of different strategies
    NetworkAPPrecisionRecallF1Time/s
    SSD74.16%76.21%66.30%0.710.028 2
    Faster-RCNN83.26%73.75%85.65%0.790.075 1
    YOLOv384.25%81.56%81.65%0.820.049 8
    YOLOv487.46%91.79%77.09%0.840.031 3
    Ours95.37%94.25%92.99%0.940.035 2
    Table 3. Comparative experimental results of different detection algorithms
    NetworkAPPrecisionRecallF1
    SSD67.59%68.53%61.51%0.65
    Faster-RCNN80.03%68.58%83.76%0.75
    YOLOv485.56%78.91%76.97%0.78
    Ours94.52%91.70%89.77%0.91
    Table 4. Comparative experimental results of different detection algorithms after adding noise
    Shuai HAO, Shan GAO, Xu MA, Beiyi AN, Tian HE, Hu WEN, Feng WANG. Infrared Pedestrian Detection Based on Cross-scale Feature Aggregation and Hierarchical Attention Mapping[J]. Acta Photonica Sinica, 2022, 51(6): 0610006
    Download Citation