Introduction
Unmanned Aerial Vehicles (UAVs) are increasingly deployed for night-time operations, yet they face significant challenges due to low-light conditions. Standard visual tracking algorithms, typically optimized for daylight, struggle with low contrast, heavy noise, and poor visibility in nocturnal environments. While existing solutions like low-light enhancement or unsupervised domain adaptation attempt to mitigate these issues, they often impose a heavy "computational tax." Transformer-based models, though accurate, are frequently too slow or memory-intensive for the constrained hardware of UAVs.
To address these challenges, we present MambaTrack, a highly efficient tracking framework designed specifically for night-time UAV operations. By replacing traditional Transformer architectures with State Space Models (Mamba), MambaTrack achieves state-of-the-art accuracy while significantly reducing computational overhead.
Methodology: Dual-Enhancement via Mamba
MambaTrack leverages the linear computational complexity of Mamba blocks to provide high performance with low latency. The architecture is built around a "dual-enhancement" strategy that combines visual and linguistic information.
The framework consists of three primary components:
- Mamba-based Low-light Enhancer (MLLE): This module integrates an illumination estimator to brighten the scene and a damage restorer to suppress noise and artifacts inherent in amplified dark signals.
- Cross-modal Mamba (CMM) Network: This module fuses visual features with natural language prompts (e.g., "a white car running on the road"). A gating mechanism dynamically selects the most relevant features from both modalities, providing a "semantic anchor" when visual data is degraded.
- Mamba Encoders: The core backbone utilizes Mamba blocks instead of standard Attention blocks, capturing long-range dependencies without the quadratic memory cost associated with Transformers.
Datasets and Benchmarking
MambaTrack was trained on large-scale datasets including TrackingNet, GOT-10k, LaSOT, COCO, and WebUAV-3M. Evaluation was conducted across five challenging night-time UAV benchmarks: DarkTrack2021, NAT2021, NAT2021L, UAVDark70, and UAVDark135.
A key contribution of this work is the manual annotation of 518 language prompts for these night-time datasets, enabling the first comprehensive vision-language tracking evaluation in low-light aerial scenarios.
Key Results and Efficiency Gains
MambaTrack sets a new benchmark for efficiency and performance in night-time tracking. By transitioning from Transformers to Mamba, the model achieves linear scaling, making sophisticated multimodal tracking viable on edge devices.
Key performance metrics include:
- Speed: Reaches 42 FPS on an RTX A6000, representing a 2.8x speedup over the previous state-of-the-art, CiteTracker.
- Memory Efficiency: Reduces GPU memory usage by 50.2% and learnable parameters by up to 91% compared to CiteTracker.
- Accuracy: Achieved a 57.2% AUC on the UAVDark135 dataset, outperforming both pure visual and vision-language competitors.
Ablation studies confirmed that both the MLLE (visual enhancement) and the CMM (language fusion) are critical; removing either component led to significant performance degradation.
Conclusion
MambaTrack demonstrates that high-performance night-time UAV tracking does not require prohibitive computational resources. By combining Mamba-based efficiency with dual visual-linguistic enhancement, the framework provides a robust solution for low-light aerial surveillance. Future work will focus on deploying MambaTrack on actual UAV onboard platforms, such as NVIDIA Jetson modules, to validate real-time performance in the field.
Summary updated