Figures (5)  Tables (3)
    • Figure 1. 

      Images of the gap between PSDs and metro doors.

    • Figure 2. 

      Schematic diagram of data collection and sample images.

    • Figure 3. 

      Network architecture[37] of YOLOv3.

    • Figure 4. 

      Building blocks of ShuffleNetV1 and ShuffleNetV2.

    • Figure 5. 

      FPS vs mAP on the present dataset. FPS is measured on the RTX 3090 GPU.

    • Foreign object categories Training and validation Testing Total
      Rope 472 123 595
      Cord 363 96 509
      Wig 19 5 24
      School bag 70 19 89
      Plastic bag 507 121 628
      Box 72 15 87
      Shoulder bag 292 66 358
      Wallet 549 136 690
      Cell phone 491 134 625
      Bottle 722 206 928
      Umbrella 79 15 94
      Person 87 10 97
      Others 53 12 65
      Normal 510 124 634
      Cardboard 362 100 512
      Total 4748 1187 5935

      Table 1. 

      The statistics of the current dataset.

    • Algorithms Backbone Size mAP@0.5 FPS Model size (MB)
      SSD VGG 300 × 300 0.859 126 97.7
      YOLOv3 Darknet-53 640 × 640 0.889 213 117
      YOLOv3 Darknet-53-SPP 640 × 640 0.879 208 119
      YOLOv4 CSPDarknet53 with Mish activation 640 × 480 0.869 92 245
      YOLOv4 Leaky-CSPDarknet53 with Leaky activation 640 × 480 0.876 93 245
      YOLOv4 SAM-Leaky-CSPDarknet53 with Leaky activation-SAM 640 × 480 0.876 86 250
      YOLOv4 Mish-CSPDarknet53 with Mish activation 640 × 480 0.867 92 245
      YOLOv4 SAM-Mish-CSPDarknet53 with Mish activation-SAM 640 × 480 0.874 86 250
      YOLOv5-m CSPDarknet-SPP 640 × 640 0.884 233 40.6
      YOLOv5-l CSPDarknet-SPP 640 × 640 0.884 154 89.5
      YOLOv5-x** CSPDarknet-SPP 640 × 640 0.894 72 167
      YOLOX-m Modified CSP in YOLOv5 640 × 640 0.865 149 194
      YOLOX-l Modified CSP in YOLOv5 640 × 640 0.868 105 364
      YOLOX-x Modified CSP in YOLOv5 640 × 640 0.854 57 757
      YOLOX-DarkNet53 Darknet-53 640 × 640 0.848 92 487
      PPYOLOv1 ResNet18-vd 512 × 512 0.831 75 49.5
      PPYOLOv1 ResNet50-vd-dcn 608 × 608 0.843 47 178
      PPYOLOv2 ResNet50-vd-dcn 640 × 640 0.849 42 207
      PPYOLOv2 ResNet101-vd-dcn 640 × 640 0.855 37 279
      SSD MobileNetV1 300 × 300 0.819 157 22
      SSDLite MobileNetV1 300 × 300 0.865 159 23
      SSDLite MobileNetV3-Small 320 × 320 0.849 140 5.1
      SSDLite MobileNetV3-Large 320 × 320 0.857 143 11
      SSDLite GhostNet 320 × 320 0.868 142 23
      YOLOv3 MobileNetV1 608 × 608 0.847 83 93
      YOLOv3 MobileNetV3 608 × 608 0.854 80 89
      YOLOv3-Tiny* Darknet-53 640 × 640 0.854 1667 16.6
      YOLOv4-Tiny CSPDarknet-53 640 × 480 0.831 549 22.5
      YOLOv5-s*** CSPDarknet-SPP 640 × 640 0.88 588 13.7
      YOLOv5-Lite ShuffleNetv2 640 × 640 0.871 1250 3.3
      YOLOX-s Modified CSP in YOLOv5 640 × 640 0.848 282 69
      YOLOX-Tiny Modified CSP in YOLOv5 640 × 640 0.854 560 39
      YOLOX-Nano Modified CSP in YOLOv5 640 × 640 0.84 804 7.3
      PPYOLOv1 MobileNetV3-Small 320 × 320 0.856 147 9.9
      PPYOLOv1 MobileNetV3-Large 320 × 320 0.865 148 18
      PPYOLOv1 PPYOLO-Tiny 320 × 320 0.818 190 3.95
      * Fastest, ** highest mAP, *** best one in comparison.

      Table 2. 

      Experimental results on the present dataset.

    • Class YOLOv5-x YOLOv5-s
      P R F1 mAP@0.5 mAP@0.5:0.95 P R F1 mAP@0.5 mAP@0.5:0.95
      Rope 0.883 0.764 0.819 0.839 0.515 0.865 0.729 0.791 0.836 0.499
      Cord 0.826 0.693 0.754 0.822 0.41 0.757 0.646 0.697 0.75 0.384
      Wig 0.779 0.8 0.789 0.938 0.521 0.789 0.8 0.794 0.84 0.426
      School bag 0.899 0.737 0.81 0.84 0.516 0.901 0.737 0.811 0.825 0.461
      Plastic bag 0.972 0.967 0.969 0.979 0.695 0.967 0.977 0.972 0.97 0.68
      Box 0.975 1 0.987 0.995 0.811 0.976 1 0.988 0.995 0.809
      Shoulder bag 0.955 0.985 0.97 0.984 0.72 0.969 0.985 0.977 0.983 0.703
      Wallet 0.874 0.887 0.88 0.912 0.452 0.903 0.924 0.913 0.906 0.479
      Cell phone 0.989 0.963 0.976 0.987 0.561 0.962 0.957 0.959 0.962 0.556
      Bottle 0.994 0.995 0.994 0.993 0.642 0.994 0.995 0.994 0.993 0.639
      Umbrella 0.865 1 0.928 0.946 0.64 0.86 0.933 0.895 0.899 0.586
      Person 0.896 1 0.945 0.995 0.942 0.964 1 0.982 0.995 0.902
      Others 0.581 0.579 0.58 0.405 0.22 0.666 0.666 0.666 0.477 0.208
      Normal 0.821 0.628 0.712 0.794 0.399 0.834 0.649 0.73 0.811 0.415
      Cardboard 0.971 0.96 0.965 0.982 0.576 0.96 0.98 0.97 0.958 0.576
      All 0.885 0.864 0.874 0.894 0.575 0.891 0.865 0.878 0.88 0.555
      P: precision, R: recall, mAP@0.5:0.95: average mAP over different IoU thresholds, from 0.5 to 0.95, step 0.05 (0.5, 0.55, 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9, 0.95).

      Table 3. 

      PR values and F1 scores of YOLOv5-x, -s.