[Paper Review] Towards Robust Visual Tracking for Unmanned Aerial Vehicle with Tri-Attentional Correlation Filters

Yujie He, Changhong Fu+3 more|arXiv (Cornell University)|Aug 2, 2020

Video Surveillance and Tracking Methods41 references4 citations

TL;DR

This paper proposes a tri-attention correlation filter (TACF) tracker for robust UAV visual tracking by integrating contextual, dimensional, and spatiotemporal attention mechanisms into a correlation filter framework. The method enhances discriminative power against partial occlusion and cluttered backgrounds while maintaining real-time performance at ~28 FPS, achieving 4.8% and 8.2% relative gains in success rate on UAVDT and UAV123@10fps benchmarks over the baseline KCC tracker.

ABSTRACT

Object tracking has been broadly applied in unmanned aerial vehicle (UAV) tasks in recent years. However, existing algorithms still face difficulties such as partial occlusion, clutter background, and other challenging visual factors. Inspired by the cutting-edge attention mechanisms, a novel object tracking framework is proposed to leverage multi-level visual attention. Three primary attention, i.e., contextual attention, dimensional attention, and spatiotemporal attention, are integrated into the training and detection stages of correlation filter-based tracking pipeline. Therefore, the proposed tracker is equipped with robust discriminative power against challenging factors while maintaining high operational efficiency in UAV scenarios. Quantitative and qualitative experiments on two well-known benchmarks with 173 challenging UAV video sequences demonstrate the effectiveness of the proposed framework. The proposed tracking algorithm favorably outperforms 12 state-of-the-art methods, yielding 4.8% relative gain in UAVDT and 8.2% relative gain in UAV123@10fps against the baseline tracker while operating at the speed of $\sim$ 28 frames per second.

Motivation & Objective

• To address the limitations of correlation filter-based trackers in UAV scenarios, particularly under partial occlusion and background clutter.
• To enhance discriminative power by integrating multi-level visual attention into the correlation filter training and detection pipeline.
• To maintain real-time performance suitable for onboard UAV systems while improving robustness against challenging visual conditions.
• To develop a unified attention framework that leverages contextual, dimensional, and spatiotemporal cues for improved response map quality.

Proposed method

• Introduces a contextual attention module that assigns dynamic penalty factors to context patches during filter training, improving environmental perception and object discrimination.
• Applies dimensional attention to focus on informative feature dimensions in the response map, reducing noise impact by suppressing less relevant features.
• Employs spatiotemporal attention to enhance response maps by leveraging temporal consistency across consecutive frames, improving localization accuracy.
• Fuses all three attention mechanisms—contextual, dimensional, and spatiotemporal—into a unified tri-attention correlation filter (TACF) framework for end-to-end tracking.
• Utilizes Fourier-domain operations to maintain computational efficiency, enabling real-time inference at ~28 FPS on a single CPU.
• Employs a learnable activation threshold βt in the dimensional attention module, optimized to 0.1 for peak performance on UAVDT.

Figure 1: Comparison between baseline KCC tracker [ 4 ] and the proposed TACF tracker. In UAV tracking, TACF incorporates the contextual information into the filter training stage, so that the attentional filters can sense the varying appearance of the object and environment simultaneously. After th

Experimental results

Research questions

• RQ1 Can integrating multi-level attention mechanisms into correlation filters improve robustness against partial occlusion in UAV tracking?
• RQ2 How does contextual attention enhance filter training by incorporating environmental context without introducing excessive noise?
• RQ3 To what extent do dimensional and spatiotemporal attention improve response map quality and tracking accuracy in complex UAV scenes?
• RQ4 Can the proposed tri-attention framework maintain real-time performance while significantly outperforming state-of-the-art trackers in accuracy and robustness?

Key findings

• The TACF tracker achieves a 4.8% relative improvement in success rate on the UAVDT benchmark compared to the baseline KCC tracker.
• On the UAV123@10fps benchmark, TACF achieves an 8.2% relative gain in success rate over the baseline, with a success rate of 0.456.
• The ablation study confirms that each attention module (CA, DA, SA) contributes independently to performance, with the full TACF model achieving the highest success rate of 0.437 on UAVDT.
• The average processing time per frame is 35.59 ms (28.1 FPS), demonstrating real-time feasibility despite added attention modules.
• The optimal activation threshold βt for dimensional attention is found to be 0.1, yielding peak performance on UAVDT.
• The TACF tracker outperforms 12 state-of-the-art trackers in both accuracy and robustness across two major UAV benchmarks.

Figure 2: The main workflow of the TACF tracker. The object and context patches near the object are first extracted and represented as training samples. Then, the contextual attention strategy is implemented to realize adaptive learning of the surrounding environment throughout the online tracking p