Optics and Precision Engineering, Volume. 31, Issue 2, 263(2023)
Gait recognition algorithm in dense occlusion scene
Figures & Tables(13)
Fig. 1. Simulation dataset of gait contour in dense occlusion scene
Fig. 2. Output of the data augmentation method based on random binary expansion
Fig. 3. Feature representation redundancy of HPP structure in case of strict spatial alignment
Table 1. Analysis of accuracy benefit of horizontal pyramid pooling with different scales
View tableView in ArticleTable 1. Analysis of accuracy benefit of horizontal pyramid pooling with different scales
网络结构 HPP分割数 0 18 36 54 72 90 108 126 144 162 180 mean NM GaitSet ［ 1 -2 ，4 ，8 ，16 ］90.8 97.9 99.4 96.9 93.6 91.7 95.0 97.8 98.9 96.8 85.8 95.0 GaitSet ［ 1 -2 ，4 ，8 ］91.4 97.8 98.6 97.1 92.9 91.0 93.9 97.1 98.3 96.9 86.8 94.7 GaitSet ［ 1 -2 ，4 ］90.1 95.6 97.6 95.9 90.8 90.4 92.1 96.5 96.8 95.2 85.2 93.3 GaitSet ［ 1 -2 ］89.2 95.9 96.9 94.5 92.1 89.1 92.0 95.8 96.2 93.4 81.4 92.4 GaitSet ［ 1 ］75.8 83.1 87.6 87.1 82.1 78.7 82.0 89.1 86.7 81.4 68.3 82.0 BG GaitSet ［ 1 -2 ，4 ，8 ，16 ］83.8 91.2 91.8 88.8 83.3 81.0 84.1 90.0 92.2 94.5 79.0 87.2 GaitSet ［ 1 -2 ，4 ，8 ］82.6 89.6 92.5 88.7 85.8 79.2 84.2 88.2 91.9 91.3 82.7 87.0 GaitSet ［ 1 -2 ，4 ］81.6 88.6 91.8 86.1 79.2 73.7 80.3 86.4 90.5 87.3 77.3 83.9 GaitSet ［ 1 -2 ］78.6 85.5 89.2 86.0 78.6 71.4 74.9 85.3 89.6 86.4 73.2 81.7 GaitSet ［ 1 ］65.3 74.2 80.0 73.0 67.9 59.4 66.8 74.4 80.6 72.7 62.4 70.6 CL GaitSet ［ 1 -2 ，4 ，8 ，16 ］61.4 75.4 80.7 77.3 72.1 70.1 71.5 73.5 73.5 68.4 50.0 70.4 GaitSet ［ 1 -2 ，4 ，8 ］62.5 70.6 75.6 72.1 64.9 63.2 65.5 67.5 67.0 64.8 52.3 66.0 GaitSet ［ 1 -2 ，4 ］59.3 69.4 74.0 71.7 64.0 63.4 65.0 66.8 67.3 62.5 46.5 64.5 GaitSet ［ 1 -2 ］51.2 61.1 66.0 63.9 64.4 63.3 62.0 64.9 63.1 56.9 45.7 60.2 GaitSet ［ 1 ］45.1 55.2 62.2 60.7 60.0 51.7 55.2 56.4 54.5 48.6 39.7 53.6
Table 2. Gait recognition accuracy of main and auxiliary channels in different scenarios
View tableView in ArticleTable 2. Gait recognition accuracy of main and auxiliary channels in different scenarios
数据集 场景 通道 0 18 36 54 72 90 108 126 144 162 180 mean CASIA_B NM 双通道 93.0 98.4 99.6 98.7 94.0 92.0 95.8 98.4 99.1 97.7 90.0 96.1 主通道 90.6 98.7 99.0 98.3 93.6 91.9 95.7 98.3 98.4 97.4 88.5 95.9 辅通道 93.2 97.6 99.6 98.5 94.1 92.3 95.8 98.5 99.1 98.3 90.6 96.5 BG 双通道 86.8 91.5 93.8 92.2 87.7 82.2 87.9 93.3 95.1 94.0 82.0 89.7 主通道 83.0 90.3 93.3 91.8 87.0 81.0 86.9 92.4 94.2 92.6 81.4 89.5 辅通道 86.6 92.1 93.8 92.0 87.0 82.5 87.1 92.9 94.4 94.1 81.3 90.3 CL 双通道 66.0 77.4 80.9 76.6 71.0 67.1 69.5 75.0 77.2 74.1 58.7 72.1 主通道 65.8 78.5 82.4 78.5 72.6 68.1 71.3 76.3 77.3 73.1 60.4 73.1 辅通道 63.3 73.8 78.7 74.0 67.8 64.7 68.1 72.1 74.9 72.3 57.5 69.7 遮挡仿真
数据集
NM 双通道 70.0 84.4 93.8 91.3 85.9 80.9 85.9 89.8 89.0 84.8 66.2 83.8 主通道 67.6 81.8 91.8 89.9 84.5 77.8 84.8 90.5 86.6 80.8 63.7 81.8 辅通道 66.7 81.1 89.4 88.5 84.9 79.4 85.1 88.2 86.9 81.5 61.7 81.2 BG 双通道 61.9 74.9 83.6 80.5 72.6 66.4 73.2 82.0 80.4 73.2 56.4 73.2 主通道 59.0 71.7 81.1 78.9 72.0 65.1 72.6 78.4 79.2 70.9 55.5 71.3 辅通道 55.3 70.6 80.3 76.7 69.5 62.6 70.0 78.0 76.9 71.1 55.2 69.7 CL 双通道 31.9 52.9 56.9 56.2 56.0 54.1 57.3 60.7 50.1 45.6 33.7 50.5 主通道 34.4 53.7 58.0 57.7 56.4 51.1 55.6 60.3 54.6 46.3 31.7 50.9 辅通道 28.8 45.4 51.7 49.7 51.9 48.5 54.0 54.9 45.5 39.5 29.6 45.4
Table 3. Performance difference between multi-channel training and single channel training
View tableView in ArticleTable 3. Performance difference between multi-channel training and single channel training
0 18 36 54 72 90 108 126 144 162 180 mean NM 主通道联合 90.6 98.7 99.0 98.3 93.6 91.9 95.7 98.3 98.4 97.4 88.5 95.5 主通道裁剪 90.0 97.5 98.8 97.9 93.1 92.1 94.8 98.2 97.8 96.0 85.8 94.7 辅通道联合 93.2 97.6 99.6 98.5 94.1 92.3 95.8 98.5 99.1 98.3 90.6 96.1 辅通道裁剪 91.1 97.5 99.0 97.7 94.1 92.3 94.1 97.7 98.1 96.8 88.3 95.2 BG 主通道联合 83.0 90.3 93.3 91.8 87.0 81.0 86.9 92.4 94.2 92.6 81.4 88.5 主通道裁剪 84.4 91.0 93.9 90.1 86.8 80.4 85.7 90.5 93.1 91.3 80.0 87.9 辅通道联合 86.6 92.1 93.8 92.0 87.0 82.5 87.1 92.9 94.4 94.1 81.3 89.4 辅通道裁剪 85.1 91.5 93.6 91.0 86.7 81.3 84.9 91.3 94.1 92.4 80.8 88.4 CL 主通道联合 65.8 78.5 82.4 78.5 72.6 68.1 71.3 76.3 77.3 73.1 60.4 73.1 主通道裁剪 65.0 76.4 81.3 79.2 73.0 68.7 69.9 77.3 76.9 70.6 55.6 72.2 辅通道联合 63.3 73.8 78.7 74.0 67.8 64.7 68.1 72.1 74.9 72.3 57.5 69.7 辅通道裁剪 57.7 71.9 77.4 71.7 65.5 67.7 67.8 70.6 73.0 66.7 50.9 67.4
Table 4. Performance comparison between the proposed algorithm and the baseline algorithm in dense occlusion simulation dataset
View tableView in ArticleTable 4. Performance comparison between the proposed algorithm and the baseline algorithm in dense occlusion simulation dataset
0 18 36 54 72 90 108 126 144 162 180 mean NM GaitSet 67.0 82.2 91.1 91.0 84.5 81.3 87.3 90.3 89.4 83.5 64.0 82.9 Our 85.1 93.6 97.5 96.4 89.5 87.1 91.3 94.8 95.3 92.1 81.1 91.3 BG GaitSet 61.9 75.6 81.8 79.4 71.0 65.4 74.8 81.1 81.3 73.2 58.8 73.1 Our 82.0 89.7 91.5 87.8 83.6 77.1 81.5 90.0 91.3 89.5 76.9 85.5 CL GaitSet 41.9 53.9 58.5 58.5 60.8 56.3 56.8 59.6 58.2 49.3 35.1 53.5 Our 56.9 70.9 74.8 73.4 64.7 61.1 65.3 69.4 69.7 68.1 52.0 66.0
Table 5. Performance comparison of the proposed algorithm and other latest algorithms on Casia-B dataset
View tableView in ArticleTable 5. Performance comparison of the proposed algorithm and other latest algorithms on Casia-B dataset
0 18 36 54 72 90 108 126 144 162 180 mean NM CNN-3D 87.1 93.2 97.0 94.6 90.2 88.3 91.1 93.8 96.5 96.0 85.7 92.1 STGAN - - - 94.6 - 88.3 - 93.8 - - - 92.2 Bi-Route 87.6 93.2 97.3 96.1 93.0 90.3 92.3 96.3 96.6 94.2 85.3 92.9 ICNet 90.8 97.9 99.4 96.9 93.6 91.7 95.0 97.8 98.9 96.8 85.8 95.0 Two-Path 88.1 96.7 97.8 96.6 92.6 90.5 93.4 96.6 97.3 95.4 85.5 93.7 GaitSet 90.8 97.9 99.4 96.9 93.6 91.7 95.0 97.8 98.9 96.8 85.8 95.0 MBRDNet - - - - - - - - - - - 94.9 our 91.8 98.3 99.3 98.0 92.4 90.9 94.2 97.3 98.1 96.8 88.4 95.0 BG CNN-LB 64.2 80.6 82.7 76.9 64.8 63.1 68.0 76.9 82.2 75.4 61.3 72.4 Bi-Route 77.5 88.8 92.5 93.6 86.1 84.6 88.3 92.3 89.8 86.6 78.7 87.2 ICNet 83.8 91.2 91.8 88.8 83.3 81.0 84.1 90.0 92.2 94.4 79.0 87.2 Two-Path 76.4 85.2 88.2 83.9 78.7 74.9 77.8 85.3 88.5 85.5 76.1 81.9 GaitSet 83.8 91.2 91.8 88.8 83.3 81.0 84.1 90.0 92.2 94.4 79.0 87.2 MBRDNet - - - - - - - - - - - 88.2 our 89.0 95.1 94.6 91.8 88.8 84.1 88.2 92.7 95.1 93.6 83.5 90.6 CL CNN-LB 37.7 57.2 66.6 61.1 55.2 54.6 55.2 59.1 58.9 48.8 39.4 54.0 Bi-Route 54.6 65.6 69.1 72.7 70.6 69.2 68.4 70.5 66.7 63.4 51.3 65.6 ICNet 61.4 75.4 80.7 77.3 72.1 70.1 71.5 73.5 73.5 68.4 50.0 70.4 Two-Path 60.2 72.0 76.1 72.6 68.8 68.1 66.3 67.6 68.8 64.4 51.8 67.0 GaitSet 61.4 75.4 80.7 77.3 72.1 70.1 71.5 73.5 73.5 68.4 50.0 70.4 MBRDNet - - - - - - - - - - - 72.1 our 67.9 81.6 84.3 81.6 74.1 72.9 74.3 75.9 77.7 76.8 60.3 75.2
Table 6. Performance comparison of the proposed algorithm and other algorithms on Casia-C dataset
View tableView in ArticleTable 6. Performance comparison of the proposed algorithm and other algorithms on Casia-C dataset
普通 背包 快走 慢走 FBC 84.6 - 83.6 88 OFCD+PCA+LDA 97 - 87 88 GaitSet 100 81.3 95 89.2 our 100 89.6 98.8 97.9
Table 7. Ablation experimental results of the proposed algorithm on dense occlusion simulation dataset
View tableView in ArticleTable 7. Ablation experimental results of the proposed algorithm on dense occlusion simulation dataset
C R D T 0 18 36 54 72 90 108 126 144 162 180 mean Dims NM GaitSet × × × × 67.0 82.2 91.1 91.0 84.5 81.3 87.3 90.3 89.4 83.5 64.0 82.9 15 872 our1 √ × × × 73.9 85.7 92.9 91.8 86.0 82.3 87.0 93.0 92.1 87.6 67.7 85.5 15 872 our2 √ √ × × 80.8 93.2 97.5 95.4 90.5 86.4 90.6 95.5 94.4 90.3 77.1 90.2 15 872 our3 √ √ √ × 84.9 94.1 97.4 95.8 91.2 89.2 91.3 95.1 95.8 94.6 82.2 92.0 8 192 our4 √ √ √ √ 85.1 93.6 97.5 96.4 89.5 87.1 91.3 94.8 95.3 92.1 81.1 91.3 4 096 BG GaitSet × × × × 61.9 75.6 81.8 79.4 71.0 65.4 74.8 81.1 81.3 73.2 58.8 73.1 15 872 our1 √ × × × 66.4 81.5 83.6 80.5 76.5 66.4 74.2 81.7 83.1 76.5 61.9 75.7 15 872 our2 √ √ × × 78.7 86.4 89.8 85.6 81.3 78.5 82.4 87.7 89.5 86.0 73.1 83.5 15 872 our3 √ √ √ × 83.2 89.5 92.4 88.3 84.3 78.5 82.4 89.6 92.0 90.8 77.3 86.2 8 192 our4 √ √ √ √ 82.0 89.7 91.5 87.8 83.6 77.1 81.5 90.0 91.3 89.5 76.9 85.5 4 096 CL GaitSet × × × × 41.9 53.9 58.5 58.5 60.8 56.3 56.8 59.6 58.2 49.3 35.1 53.5 15 872 our1 √ × × × 38.7 55.6 61.2 59.4 60.4 53.1 55.8 60.8 58.6 50.1 31.6 53.2 15 872 our2 √ √ × × 53.1 68.8 71.5 70.9 65.3 62.5 64.2 68.4 67.5 59.4 43.9 63.2 15 872 our3 √ √ √ × 55.0 70.6 72.4 71.0 65.6 62.5 65.5 69.9 69.7 65.6 49.6 65.2 8 192 our4 √ √ √ √ 56.9 70.9 74.8 73.4 64.7 61.1 65.3 69.4 69.7 68.1 52.0 66.0 4 096
Table 8. Ablation experimental results of the proposed algorithm on Casia-B dataset
View tableView in ArticleTable 8. Ablation experimental results of the proposed algorithm on Casia-B dataset
C R D T 0 18 36 54 72 90 108 126 144 162 180 mean Dims NM GaitSet × × × × 90.8 97.9 99.4 96.9 93.6 91.7 95.0 97.8 98.9 96.8 85.8 95.0 15 872 our1 √ × × × 90.3 98.1 98.9 98.3 92.4 89.5 93.6 97.4 98.3 96.5 85.7 94.5 15 872 our2 √ √ × × 90.4 97.2 98.5 96.8 92.3 89.3 93.2 96.4 97.2 96.0 85.5 93.9 15 872 our3 √ √ √ × 91.9 97.8 99.4 97.9 92.7 91.3 94.8 98.1 98.2 97.4 86.7 95.1 8 192 our4 √ √ √ √ 91.8 98.3 99.3 98.0 92.4 90.9 94.2 97.3 98.1 96.8 88.4 95.0 4 096 BG GaitSet × × × × 83.8 91.2 91.8 88.8 83.3 81.0 84.1 90.0 92.2 94.5 79.0 87.2 15 872 our1 √ × × × 84.8 91.9 93.0 89.1 85.9 78.6 83.2 90.9 93.9 90.4 79.0 87.3 15 872 our2 √ √ × × 88.1 92.8 93.8 90.7 86.2 81.9 85.6 91.5 93.2 92.9 81.8 89.0 15 872 our3 √ √ √ × 89.0 95.7 94.7 91.8 89.1 84.7 87.7 92.6 95.2 93.6 84.4 90.8 8 192 our4 √ √ √ √ 89.0 95.1 94.6 91.8 88.8 84.1 88.2 92.7 95.1 93.6 83.5 90.6 4 096 CL GaitSet × × × × 61.4 75.4 80.7 77.3 72.1 70.1 71.5 73.5 73.5 68.4 50.0 70.4 15 872 our1 √ × × × 66.1 78.5 80.4 76.8 71.1 70.1 71.4 74.2 75.1 71.7 57.6 72.1 15 872 our2 √ √ × × 67.4 78.9 83.0 80.2 74.0 69.5 71.9 75.5 75.3 71.8 58.9 73.3 15 872 our3 √ √ √ × 67.5 82.2 85.4 82.7 74.4 70.9 75.5 77.9 78.1 77.6 63.1 75.9 8 192 our4 √ √ √ √ 67.9 81.6 84.3 81.6 74.1 72.9 74.3 75.9 77.7 76.8 60.3 75.2 4 096
Tools
Get Citation
Copy Citation Text
Yi GAO, Miao HE. Gait recognition algorithm in dense occlusion scene[J]. Optics and Precision Engineering, 2023, 31(2): 263
Paper Information
Category: Information Sciences
Received: Sep. 11, 2022
Accepted: --
Published Online: Feb. 9, 2023
The Author Email: HE Miao (hemiao@sia.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20