Abstract
Single Object Tracking (SOT) on Unmanned Aerial Vehicles (UAVs) requires a delicate balance between computational efficiency and robustness against environmental distractors. We present BOBBY2, a lightweight Siamese-based tracker optimized for resource-constrained hardware. By implementing a novel "exemplar buffer" to maintain target appearance history and utilizing a streamlined AlexNet backbone, BOBBY2 achieves a processing speed of 85 FPS while significantly improving robustness. Our results on the GOT-10k benchmark demonstrate superior performance compared to standard Siamese trackers, maintaining high accuracy even under aggressive frame-dropping regimes.
Introduction
Autonomous drone navigation relies heavily on real-time SOT. However, existing solutions often fall into two extremes: fast but fragile correlation filters, or accurate but computationally intensive deep learning models. Modern Siamese trackers, while robust, typically exceed the power and processing budgets of small multi-rotor UAVs. BOBBY2 addresses this "Goldilocks" challenge by providing a robust tracking framework that is lean enough for embedded edge deployment without sacrificing the ability to handle target deformation and occlusions.
Methodology
BOBBY2 is built upon a Siamese convolutional neural network architecture with several key optimizations for edge performance.
Lightweight Backbone and Training
The architecture utilizes a stripped-down version of AlexNet, featuring 63% fewer parameters than the standard version. This reduction to 22.23M parameters allows for higher throughput on embedded devices. The model is trained using a "one-cycle" super-convergence regime, reaching peak performance in under two epochs. To mitigate "ghosting"—where the tracker locks onto background objects—the training set (ImageNet-VID) was augmented with over 378,000 hard negative samples.
Exemplar Buffer Module
The core innovation of BOBBY2 is the Exemplar Buffer Module. Unlike traditional trackers that rely on a single initial template, BOBBY2 maintains a buffer (size $\beta=4$) that sparsely caches the target’s appearance over time. This allows the model to adapt to rotations and shape changes as the flight progresses.
Efficiency Mechanisms
To further reduce power consumption and maintain real-time responsiveness, the system employs sparse activation and heuristic frame-dropping. The exemplar branch updates only every $\xi$ frames, using cached features in the interim. Interestingly, the authors noted that 93% of the total runtime was consumed by image cropping and resizing, highlighting that for edge computing, data pre-processing is often the primary hurdle rather than the neural network inference itself.
Experimental Results
BOBBY2 demonstrates exceptional speed, running at 85 FPS in a full pipeline and reaching up to 700 FPS as a standalone model on high-end hardware. On the GOT-10k benchmark, it achieved an Average Overlap (AO) of 0.392, outperforming SiamFC by 5.3%.
The buffer-based approach proved highly resilient; success rates dropped by only ~1% even when the buffer was intentionally populated with 75% distractor data, confirming that a single valid exemplar is often sufficient for maintenance. As shown in the performance metrics, the tracker maintains consistent scores even as the dropping interval $\eta$ increases.
Conclusion
BOBBY2 provides a viable path for high-speed, robust tracking on low-power UAVs. While it performs well on standard benchmarks, a performance gap remains when applied to high-angle, shaky UAV-specific datasets like the custom TU-3 dataset. Future work will focus on domain-specific fine-tuning and optimizing the pre-processing pipeline to alleviate the primary computational bottleneck.
Summary updated