Chinese Journal of Liquid Crystals and Displays, Volume. 39, Issue 9, 1233(2024)

Visual place recognition method based on parallel omni-dimensional dynamic attention mechanism

Peijin LIU1, Shujie LIU1, Lin HE2、*, Lijun PENG1, and Xuefeng FU1
Author Affiliations
  • 1School of Mechanical and Electrical Engineering,Xi'an University of Architecture & Technology,Xi'an 710055,China
  • 2Faculty of Science,Xi'an University of Architecture & Technology,Xi'an 710055,China
  • show less
    Figures & Tables(12)
    Flowchart of visual place recognition based on POD attention mechanism
    POD attention mechanism architecture
    Multi-head attention module
    SSE module
    Comparison of Grad-CAM experiments under different environmental changes
    • Table 1. Experimental comparison results to validate the effectiveness of POD under the Nordland dataset

      View table
      View in Article

      Table 1. Experimental comparison results to validate the effectiveness of POD under the Nordland dataset

      方法Recall@1Recall@5Recall@10

      NetVLAD

      NetVLAD&POD

      Patch-NetVLAD

      Patch-NetVLAD&POD

      13.0

      18.1

      44.9

      54.6

      20.6

      33.9

      50.2

      64.4

      25.0

      41.7

      52.2

      68.0

    • Table 2. Comparison results for each network under the nordland dataset

      View table
      View in Article

      Table 2. Comparison results for each network under the nordland dataset

      方法Recall@1Recall@5Recall@10

      Conv-AP

      DELG global

      DELG local

      HEAPUtil

      NetVLAD&GPM

      Patch-NetVLAD&POD

      38.2

      23.4

      60.1

      29.3

      38.5

      54.6

      54.8

      35.4

      63.5

      53.9

      64.4

      61.2

      41.7

      64.6

      60.7

      68.0

    • Table 3. Comparison results of each algorithms under the mapillary dataset

      View table
      View in Article

      Table 3. Comparison results of each algorithms under the mapillary dataset

      方法Recall@1Recall@5Recall@10

      NetVLAD

      NetVLAD&POD

      Patch-NetVLAD

      Patch-NetVLAD&POD

      58.6

      63.6

      76.3

      78.1

      71.2

      76.0

      79.8

      81.9

      76.1

      80.2

      86.0

      87.1

    • Table 4. Comparison of indicators before and after the addition of POD to Resnet18 and Mobilenet v3

      View table
      View in Article

      Table 4. Comparison of indicators before and after the addition of POD to Resnet18 and Mobilenet v3

      方法

      Recall

      @1

      Recall

      @5

      Recall

      @10

      Recall

      @20

      Mobilenetv3

      Mobilenetv3&POD

      Resnet18

      Resnet18&POD

      45.8

      46.9

      46.7

      52.9

      59.9

      62.0

      62.4

      69.0

      67.4

      69.8

      70.3

      74.4

      73.1

      76.5

      77.7

      78.5

    • Table 5. Running time testing experiment of improved visual place recognition algorithm

      View table
      View in Article

      Table 5. Running time testing experiment of improved visual place recognition algorithm

      测试配置

      VGG16&

      POD

      Resnet18&

      POD

      Mobilenetv3&POD

      GPU

      机器人

      0.05

      0.88

      0.03

      0.41

      0.02

      0.35

    • Table 6. Comparison of experiments with the addition of different attention mechanisms

      View table
      View in Article

      Table 6. Comparison of experiments with the addition of different attention mechanisms

      方法Recall@1Recall@5Recall@10

      Resnet18

      Resnet18&SE

      Resnet18&CBAM

      Resnet18&POD

      46.7

      49.6

      50.0

      52.9

      62.4

      63.6

      64.0

      69.0

      70.3

      71.1

      71.9

      74.4

    • Table 7. Effect of different blocks in the POD attention

      View table
      View in Article

      Table 7. Effect of different blocks in the POD attention

      方法Recall@1Recall@5Recall@10

      Resnet18

      Resnet18&SSE

      Resnet18&OD

      Resnet18&POD

      46.7

      47.1

      50.4

      52.9

      62.4

      64.0

      66.5

      69.0

      70.3

      71.1

      72.4

      74.4

    Tools

    Get Citation

    Copy Citation Text

    Peijin LIU, Shujie LIU, Lin HE, Lijun PENG, Xuefeng FU. Visual place recognition method based on parallel omni-dimensional dynamic attention mechanism[J]. Chinese Journal of Liquid Crystals and Displays, 2024, 39(9): 1233

    Download Citation

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

    Category:

    Received: Oct. 16, 2023

    Accepted: --

    Published Online: Nov. 13, 2024

    The Author Email: Lin HE (helin716@xauat.edu.cn)

    DOI:10.37188/CJLCD.2023-0328

    Topics