Infrared and Laser Engineering, Volume. 54, Issue 3, 20240496(2025)
Adaptive tracking method for infrared small targets in dynamic and complex scenes (invited)
Figures & Tables(16)
Fig. 1. Tracking framework of the proposed network (DTFE module represent dynamic template feature enhancement module; MSA module represent multi-layer self-attention module; ATU module represent adaptive template update module)
Fig. 3. Multi-layer Self-attention (MSA) module (This module consists of encoder-decoder self-attention module and pixel-level self-attention module connected in series)
Fig. 7. Visualization results of different algorithms on Seq1-Seq8
Fig. 8. Feature visualization under scale changes （The scale gradually decreases from left to right）
Fig. 10. The visualization results of the proposed method in scenarios with scale and attitude changes (Scale change in the first row, posture change in the second row)
Table 1. The structure of multi-scale feature extraction and fusion network
View tableView in ArticleTable 1. The structure of multi-scale feature extraction and fusion network
Level Filter Template Search Input - 127×127×3 255×255×3 Conv0 3×3 127×127×12 255×255×12 Basic residual 3×3 127×127×12 255×255×12 Basic residual 3×3 127×127×12 255×255×12 Maxpooling 2×2 63×63×12 127×127×12 Conv1 3×3 63×63×40 127×127×40 Maxpooling 2×2 31×31×40 63×63×40 Conv2 3×3 31×31×64 63×63×64 Maxpooling 2×2 15×15×64 31×31×64 Conv3 3×3 15×15×128 31×31×128 ASPP - 15×15×128 31×31×128 Feature fusion - 15×15×64 31×31×64
Table 2. Quantitative comparison results (success rate)
View tableView in ArticleTable 2. Quantitative comparison results (success rate)
Algorithms Seq1 Seq2 Seq3 Seq4 Seq5 Seq6 Seq7 Seq8 Speed/frame·s–1 MOSSE 0.030 0.132 0.088 0.067 0.031 0.788 0.108 0.540 580 CSK 0.052 0.008 0.180 0.038 0.068 0.434 0.082 0.021 430 BACF 0.013 0.051 0.315 0.024 0.030 0.075 0.134 0.633 83 DSST 0.053 0.051 0.322 0.021 0.031 0.461 0.165 0.615 145 KCF 0.013 0.311 0.324 0.023 0.032 0.404 0.464 0.019 502 ECO 0.810 0.132 0.335 0.021 0.038 0.797 0.137 0.738 90 SiamBAN 0.035 0.649 0.411 0.172 0.048 0.915 0.796 0.692 64 SiamCAR 0.033 0.764 0.429 0.038 0.274 0.800 0.200 0.624 33 SiamGAT 0.025 0.877 0.414 0.026 0.412 0.502 0.790 0.886 42 SiamSA 0.439 0.243 0.005 0.134 0.042 0.593 0.078 0.805 39 SmallTrack 0.024 0.760 0.337 0.074 0.343 0.443 0.689 0.569 588 Ours 0.844 0.982 0.940 0.439 0.986 0.828 1.000 0.824 105
Table 3. Quantitative comparison results (precision)
View tableView in ArticleTable 3. Quantitative comparison results (precision)
Algorithms Seq1 Seq2 Seq3 Seq4 Seq5 Seq6 Seq7 Seq8 Speed/frame·s–1 MOSSE 0.029 0.143 0.370 0.073 0.061 0.853 0.371 0.883 580 CSK 0.056 0.014 0.464 0.030 0.108 0.458 0.242 0.069 430 BACF 0.014 0.084 0.462 0.029 0.060 0.089 0.293 0.915 83 DSST 0.058 0.086 0.453 0.028 0.062 0.515 0.315 0.907 145 KCF 0.014 0.405 0.465 0.029 0.071 0.475 0.673 0.068 502 ECO 0.918 0.140 0.461 0.028 0.068 0.865 0.291 0.940 90 SiamBAN 0.035 0.669 0.424 0.161 0.052 0.915 0.814 0.948 64 SiamCAR 0.033 0.907 0.469 0.062 0.622 0.825 0.384 1.000 33 SiamGAT 0.025 0.920 0.421 0.069 0.522 0.472 1.000 0.967 42 SiamSA 0.416 0.286 0.017 0.149 0.042 0.625 0.338 1.000 39 SmallTrack 0.037 0.949 0.565 0.094 0.577 0.596 0.956 0.969 588 Ours 0.855 0.928 0.993 0.604 0.994 0.897 1.000 0.995 105
Table 4. The success rate (
IOU ≥0.5) and precision (P ≤ 5 pixel) of different backbone networksView tableView in ArticleTable 4. The success rate (
IOU ≥0.5) and precision (P ≤ 5 pixel) of different backbone networksAverage SR Average PRE Alexnet 0.317 0.495 Resnet18 0.575 0.750 Resnet50 0.449 0.596 Our backbone 0.702 0.782
Table 5. The effectiveness of each proposed module in improving tracking performance
View tableView in ArticleTable 5. The effectiveness of each proposed module in improving tracking performance
DTFE MSA ATU Average SR Average PRE Baseline 0.702 0.782 √ 0.749 0.823 √ 0.730 0.814 √ 0.762 0.831 √ √ 0.804 0.862 √ √ 0.807 0.862 √ √ 0.803 0.868 √ √ √ 0.855 0.915
Table 6. Quantitative analysis of performance indicators under scale and attitude changes
View tableView in ArticleTable 6. Quantitative analysis of performance indicators under scale and attitude changes
Scene Average SR Average PRE Scale change 0.851 0.902 Posture change 0.890 0.911
Tools
Get Citation
Copy Citation Text
Tianlei MA, Xinhao LIU, Jinzhu PENG, Zhiqiang KAI, Hao WANG. Adaptive tracking method for infrared small targets in dynamic and complex scenes (invited)[J]. Infrared and Laser Engineering, 2025, 54(3): 20240496
Paper Information
Category: Optical imaging, display and information processing
Received: Oct. 31, 2024
Accepted: --
Published Online: Apr. 8, 2025
The Author Email: Jinzhu PENG (jzpeng@zzu.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20