Hyperspectral Target Tracking via Spatial-Spectral Attention Weight Variance Gradient and Depth Contrast Enhancement

Area of Science:

• Computer Vision
• Remote Sensing
• Signal Processing

Background:

• Hyperspectral target tracking faces significant challenges due to scale variations.
• Existing methods struggle to maintain tracking accuracy when target scales change dynamically.
• Robust appearance modeling is crucial for effective tracking in complex environments.

Purpose of the Study:

• To propose a novel hyperspectral target tracking method robust to scale variations.
• To enhance tracking performance by integrating spatial-spectral attention mechanisms with depth estimation.
• To develop an adaptive fusion strategy for improved tracking accuracy.

Main Methods:

• Utilized spatial-spectral attention weight variance gradient for dimensionality reduction and fused attention weights.
• Implemented a dual-path preprocessing module and a Vision Transformer encoder with depth contrast enhancement.
• Employed weight adaptive mixed fusion to combine attention weights and depth information.
• Incorporated depth-aware geometric constraints and spectral-spatial information for appearance modeling.

Main Results:

• Achieved state-of-the-art performance on hyperspectral video sequences.
• Demonstrated superior robustness to scale variations compared to existing methods.
• Attained an Area Under the Curve (AUC) of 0.6704 and a Detection Precision at 20% False Positives (DP@20) of 0.9455.
• Outperformed state-of-the-art methods by 3.1% in scale variation robustness.

Conclusions:

• The proposed method effectively addresses scale variations in hyperspectral target tracking.
• The integration of spatial-spectral attention and depth estimation significantly enhances tracking robustness.
• The depth-aware approach provides a promising direction for future hyperspectral tracking research.