• Infrared and Laser Engineering
  • Vol. 54, Issue 3, 20240496 (2025)
Tianlei MA1,2, Xinhao LIU1, Jinzhu PENG1,2,*, Zhiqiang KAI1, and Hao WANG1
Author Affiliations
  • 1School of Electrical and Information Engineering, Zhengzhou University, Zhengzhou 450001, China
  • 2The State Key Laboratory of Intelligent Agricultural Power Equipment, Zhengzhou University, Luoyang 471039, China
  • show less
    DOI: 10.3788/IRLA20240496 Cite this Article
    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 Copy Citation Text show less
    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. 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)
    Dynamic Template Feature Enhancement (DTFE) module
    Fig. 2. Dynamic Template Feature Enhancement (DTFE) module
    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. 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)
    Visualization of template features
    Fig. 4. Visualization of template features
    Success rate curves of different algorithms on Seq1-Seq8
    Fig. 5. Success rate curves of different algorithms on Seq1-Seq8
    Precision curves of different algorithms on Seq1-Seq8
    Fig. 6. Precision curves of different algorithms on Seq1-Seq8
    Visualization results of different algorithms on Seq1-Seq8
    Fig. 7. Visualization results of different algorithms on Seq1-Seq8
    Feature visualization under scale changes (The scale gradually decreases from left to right)
    Fig. 8. Feature visualization under scale changes (The scale gradually decreases from left to right)
    Feature visualization under posture changes
    Fig. 9. Feature visualization under posture changes
    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)
    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)
    LevelFilterTemplateSearch
    Input-127×127×3255×255×3
    Conv03×3127×127×12255×255×12
    Basic residual3×3127×127×12255×255×12
    Basic residual3×3127×127×12255×255×12
    Maxpooling2×263×63×12127×127×12
    Conv13×363×63×40127×127×40
    Maxpooling2×231×31×4063×63×40
    Conv23×331×31×6463×63×64
    Maxpooling2×215×15×6431×31×64
    Conv33×315×15×12831×31×128
    ASPP-15×15×12831×31×128
    Feature fusion-15×15×6431×31×64
    Table 1. The structure of multi-scale feature extraction and fusion network
    AlgorithmsSeq1Seq2Seq3Seq4Seq5Seq6Seq7Seq8Speed/frame·s–1
    MOSSE0.0300.1320.0880.0670.0310.7880.1080.540580
    CSK0.0520.0080.1800.0380.0680.4340.0820.021430
    BACF0.0130.0510.3150.0240.0300.0750.1340.63383
    DSST0.0530.0510.3220.0210.0310.4610.1650.615145
    KCF0.0130.3110.3240.0230.0320.4040.4640.019502
    ECO0.8100.1320.3350.0210.0380.7970.1370.73890
    SiamBAN0.0350.6490.4110.1720.0480.9150.7960.69264
    SiamCAR0.0330.7640.4290.0380.2740.8000.2000.62433
    SiamGAT0.0250.8770.4140.0260.4120.5020.7900.88642
    SiamSA0.4390.2430.0050.1340.0420.5930.0780.80539
    SmallTrack0.0240.7600.3370.0740.3430.4430.6890.569588
    Ours0.8440.9820.9400.4390.9860.8281.0000.824105
    Table 2. Quantitative comparison results (success rate)
    AlgorithmsSeq1Seq2Seq3Seq4Seq5Seq6Seq7Seq8Speed/frame·s–1
    MOSSE0.0290.1430.3700.0730.0610.8530.3710.883580
    CSK0.0560.0140.4640.0300.1080.4580.2420.069430
    BACF0.0140.0840.4620.0290.0600.0890.2930.91583
    DSST0.0580.0860.4530.0280.0620.5150.3150.907145
    KCF0.0140.4050.4650.0290.0710.4750.6730.068502
    ECO0.9180.1400.4610.0280.0680.8650.2910.94090
    SiamBAN0.0350.6690.4240.1610.0520.9150.8140.94864
    SiamCAR0.0330.9070.4690.0620.6220.8250.3841.00033
    SiamGAT0.0250.9200.4210.0690.5220.4721.0000.96742
    SiamSA0.4160.2860.0170.1490.0420.6250.3381.00039
    SmallTrack0.0370.9490.5650.0940.5770.5960.9560.969588
    Ours0.8550.9280.9930.6040.9940.8971.0000.995105
    Table 3. Quantitative comparison results (precision)
    Average SRAverage PRE
    Alexnet0.3170.495
    Resnet180.5750.750
    Resnet500.4490.596
    Our backbone0.7020.782
    Table 4. The success rate (IOU≥0.5) and precision (P ≤ 5 pixel) of different backbone networks
    DTFEMSAATUAverage SRAverage PRE
    Baseline0.7020.782
    0.7490.823
    0.7300.814
    0.7620.831
    0.8040.862
    0.8070.862
    0.8030.868
    0.8550.915
    Table 5. The effectiveness of each proposed module in improving tracking performance
    SceneAverage SRAverage PRE
    Scale change0.8510.902
    Posture change0.8900.911
    Table 6. Quantitative analysis of performance indicators under scale and attitude changes
    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
    Download Citation