融合黑暗感知提示的夜间目标跟踪框架

Key Points

• The study aims to address feature degradation and background noise issues in nighttime target tracking.
• Developed a cross-layer consistency prompt generator (CTCP) for recovering target structures in feature space.
• Designed a prompt semantic calibration attention (PSCA) to improve attention distribution and suppress interference.
• Introduced an energy-aware dual-gating update strategy (EDGU) for reliable dynamic template updates.
• Achieved success rates of 0.557, 0.585, and 0.608 on NAT2021, LLOT, and UAVDark135 datasets respectively.
• Improved tracking performance by 3.1% over a similar method (DCPT) on NAT2021.
• Reached an inference speed of 76.8 FPS on NVIDIA A100 hardware, with only 7.03% parameter count of full fine-tuning.

Abstract