Journal of Electronic Science and Technology, Volume. 22, Issue 1, 100243(2024)

Benchmarking YOLOv5 models for improved human detection in search and rescue missions

Namat Bachir and Qurban Ali Memon*
Author Affiliations
  • Electrical Engineering Department, College of Engineering, United Arab Emirates University, Al Ain, 15551, United Arab Emirates
  • show less
    Figures & Tables(14)
    Proposed approach using YOLOv5.
    Sample dataset images.
    Input and network image resolution: (a) input image resolution and (b) network image resolution.
    Classifier at different IOU thresholds: (a) precision and (b) recall.
    Average precision: (a) precision–to-recall and (b) F1score.
    Two samples of positive detections with high confidence.
    False detections of YOLOv5L model (shadows, dark areas).
    • Table 1. Network resolution (%) YOLOv5 detection performance.

      View table
      View in Article

      Table 1. Network resolution (%) YOLOv5 detection performance.

      Network resolutiontestAPAP50
      320×3203878
      640×6406095
      832×8326496
    • Table 2. Hardware machine specification of Google Collaboratory.

      View table
      View in Article

      Table 2. Hardware machine specification of Google Collaboratory.

      ParameterSpecification
      Model nameIntel(R) Xeon(R) CPU @ 2.30 GHz
      CPU MHz2299.998
      Cache size46080 KB
      CPU cores2
      RAM12 GB
      GPUNvidia K80 / T4
      GPU memory clock0.82 GHz / 1.59 GHz
      Max execution time12 hours
      Max idle time90 minutes
      Model nameIntel(R) Xeon(R) CPU @ 2.30 GHz
      CPU (MHz)2299.998
    • Table 3. Comparative results on the SAR dataset.

      View table
      View in Article

      Table 3. Comparative results on the SAR dataset.

      ModelClassImagesLabelsmAP@0.5mAP @0.75mAP
      YOLOv5LAll79226050.9690.7430.643
      YOLOv4All79226050.9600.7100.610
      YOLOv3All79226050.9250.630.902
      Faster R-CNNAll79226050.9100.5100.500
    • Table 4. Precision-to-recall ratios for different models.

      View table
      View in Article

      Table 4. Precision-to-recall ratios for different models.

      ModelPrecisionRecall
      YOLOv5L0.9710.932
      YOLOv40.9600.910
      YOLOv30.9620.892
      Faster R-CNN0.6700.936
    • Table 5. Performance comparison of YOLOv5 models on the SAR dataset.

      View table
      View in Article

      Table 5. Performance comparison of YOLOv5 models on the SAR dataset.

      ModelPrecisionRecallmAP
      YOLOv5s0.9400.9170.933
      YOLOv5m0.7750.7670.762
      YOLOv5L0.9710.9320.969
    • Table 6. Performance comparison of YOLOv5 models on the HERIDAL dataset.

      View table
      View in Article

      Table 6. Performance comparison of YOLOv5 models on the HERIDAL dataset.

      ModelPrecisionRecallmAP
      YOLOv5s0.7530.6940.731
      YOLOv5m0.7970.8120.810
      YOLOv5L0.9000.8930.864
    • Table 7. Comparison of YOLOv3, YOLOv4, YOLOv5L and faster R-CNN models on the HERIDAL dataset.

      View table
      View in Article

      Table 7. Comparison of YOLOv3, YOLOv4, YOLOv5L and faster R-CNN models on the HERIDAL dataset.

      ModelPrecisionRecallmAP
      YOLOv5L0.9000.8930.864
      YOLOv40.8600.8300.783
      YOLOv30.8770.7800.752
      Faster R-CNN0.7700.8900.861
    Tools

    Get Citation

    Copy Citation Text

    Namat Bachir, Qurban Ali Memon. Benchmarking YOLOv5 models for improved human detection in search and rescue missions[J]. Journal of Electronic Science and Technology, 2024, 22(1): 100243

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category:

    Received: Jan. 19, 2023

    Accepted: Feb. 4, 2024

    Published Online: Jul. 5, 2024

    The Author Email: Memon Qurban Ali (qurban.memon@uaeu.ac.ae)

    DOI:10.1016/j.jnlest.2024.100243

    Topics