Infrared and Laser Engineering, Volume. 51, Issue 9, 20210924(2022)
Semantic enhanced guide feature reconstruction for occluded pedestrian detection
Figures & Tables(13)
Fig. 2. (a) Semantic feature enhancement structure diagram; (b) Improved global context bock
Fig. 4. Comparison of original image and different segmentation images. (a) Original image; (b) Fine segmentation map; (c) Rough segmentation map
Fig. 5. Comparison of feature visualization of Conv5_3 layer before and after adding each module. (a) Original image; (b) Baseline; (c) SFEM; (d) AFRM; (e) Proposed method
Fig. 6. Comparison of some test results. (a) CSP test results; (b) Proposed method test results
Table 1. Standards for dividing data set subsets
View tableView in ArticleTable 1. Standards for dividing data set subsets
Type Bare Reasonable Heavy Small All High h>50 h>50 h>50 50<h<75 h>20 Vis v>90% v>65% 0 v>65% v>20%
Table 2. Comparison of AFRM fusion methods
View tableView in ArticleTable 2. Comparison of AFRM fusion methods
Baseline Cat Sum Multiply Reasonable Heavy √ 11.00% 49.30% √ √ 10.60% 48.74% √ √ 11.20% 49.10% √ √ 10.26% 48.21%
Table 3. Comparison experiment of ablation of each module
View tableView in ArticleTable 3. Comparison experiment of ablation of each module
Baseline SFEM AFRM Reasonable Heavy √ 11.00% 49.30% √ √ 10.82% 48.77% √ √ 10.26% 48.21% √ √ √ 9.85% 47.32%
Table 4. Comparison results of different NMS types at different thresholds
View tableView in ArticleTable 4. Comparison results of different NMS types at different thresholds
Type IoU=0.5 IoU=0.6 IoU=0.7 Baseline （NMS） 49.30 49.91 53.47 Proposed method (NMS) 47.32 47.85 51.00 Proposed method (DIOU-NMS) 47.28 47.85 50.75
Table 5. Comparison of CityPersons data sets
View tableView in ArticleTable 5. Comparison of CityPersons data sets
Method Reasonable Bare Heavy Small All Time RepLoss 13.20 7.60 56.90 42.60 44.45 - OR-CNN 12.80 6.70 55.70 42.30 42.32 - PRNet 10.80 6.80 53.30 - - - PEN 10.40 7.00 47.40 - - 0.36 MSAF 9.50 7.10 48.40 15.50 - - IDC 10.70 - 50.60 14.70 41.40 - R2NMS 11.10 - 53.30 - - - CSANet 12.00 7.30 51.30 - - 0.32 APD 10.60 7.10 49.80 15.70 - 0.16 Couple 12.30 - 49.81 38.31 40.39 - CSP 11.00 7.03 49.30 16.00 - 0.33 Proposed 9.85 6.82 47.28 13.93 36.65 0.36
Table 6. Comparison of caltech data set
View tableView in ArticleTable 6. Comparison of caltech data set
Method Backbone Reasonable Heavy All ATT-part VGG-16 10.30 45.18 - RepLoss ResNet-50 5.00 47.90 59.00 OR-CNN ResNet-50 4.10 45.00 - SSNet ResNet-50 6.30 - - PAMS_FCN ResNet-50 4.50 47.40 53.70 Bi-Box VGG-16 7.61 44.40 - AR-Ped ResNet-50 4.36 48.80 - CSP ResNet-50 4.54 45.80 56.94 Proposed ResNet-50 4.32 44.04 56.34
Tools
Get Citation
Copy Citation Text
Xudan Sun, Qing Wu, Chunyan Zhao, Mandun Zhang. Semantic enhanced guide feature reconstruction for occluded pedestrian detection[J]. Infrared and Laser Engineering, 2022, 51(9): 20210924
Paper Information
Category: Image processing
Received: Nov. 30, 2021
Accepted: --
Published Online: Jan. 6, 2023
The Author Email:
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20