Learning Aberrance Repressed Correlation Filters for Real-Time UAV Tracking (1908.02231v2)

Abstract: Traditional framework of discriminative correlation filters (DCF) is often subject to undesired boundary effects. Several approaches to enlarge search regions have been already proposed in the past years to make up for this shortcoming. However, with excessive background information, more background noises are also introduced and the discriminative filter is prone to learn from the ambiance rather than the object. This situation, along with appearance changes of objects caused by full/partial occlusion, illumination variation, and other reasons has made it more likely to have aberrances in the detection process, which could substantially degrade the credibility of its result. Therefore, in this work, a novel approach to repress the aberrances happening during the detection process is proposed, i.e., aberrance repressed correlation filter (ARCF). By enforcing restriction to the rate of alteration in response maps generated in the detection phase, the ARCF tracker can evidently suppress aberrances and is thus more robust and accurate to track objects. Considerable experiments are conducted on different UAV datasets to perform object tracking from an aerial view, i.e., UAV123, UAVDT, and DTB70, with 243 challenging image sequences containing over 90K frames to verify the performance of the ARCF tracker and it has proven itself to have outperformed other 20 state-of-the-art trackers based on DCF and deep-based frameworks with sufficient speed for real-time applications.

• The paper presents ARCF, a novel method that suppresses aberrances during training to enhance UAV tracking robustness.
• It integrates a regularization term and ADMM optimization to reduce response map variations and mitigate boundary effects.
• Experimental results on UAV datasets demonstrate ARCF's superiority in precision, success rate, and real-time performance over state-of-the-art methods.

Overview of the Paper: Learning Aberrance Repressed Correlation Filters for Real-Time UAV Tracking

The paper presents a novel approach in the domain of visual tracking, particularly aimed at improving the robustness and accuracy of object tracking in unmanned aerial vehicle (UAV) applications. It introduces the concept of Aberrance Repressed Correlation Filter (ARCF), addressing specific challenges inherent to real-time UAV tracking—most notably, the boundary effects and aberrances in response maps that impair traditional Discriminative Correlation Filter (DCF) frameworks.

Problem Understanding and Methodology

Visual object tracking from UAVs involves several challenges, like occlusion, illumination changes, and fast motion dynamics. The DCF framework, known for its computational efficiency through the use of circulant matrices, suffers from boundary effects due to its limited search region, introducing noise from excessive background information. The authors argue that this noise can lead to aberrances—sudden unwanted changes in response maps—without proper management, degrading the tracker’s performance.

ARCF addresses these issues by enforcing conditions directly in the correlation filter's training phase to suppress aberrances. The proposed method integrates a regularization term to limit variations in response maps during the detection process. This technique is said to enhance the tracker's robustness, enabling it to distinguish the tracked object more effectively amidst challenging variables such as occlusion. By modulating the response map alteration rate, ARCF aims to mitigate the learning of irrelevant background features, maintaining focus on the object of interest.

Implementation and Results

The ARCF algorithm introduces a novel formulation by constructing a loss function designed to curb aberrances preemptively. The authors employ ADMM optimization to solve the minimization problem efficiently. The approach is validated through extensive experiments conducted on widely-used UAV datasets: UAV123@10fps, UAVDT, and DTB70.

Results from these experiments demonstrate ARCF's superiority over other well-recognized state-of-the-art tracking methods. The ARCF tracker not only outperforms other techniques in terms of precision and success rate but also achieves sufficient computational speed consistent with real-time tracking requirements.

Significantly, ARCF exhibits qualitative robustness, better maintaining tracking performance under conditions that typically induce aberrances, such as partial occlusions. Quantitative analysis shows a substantial reduction in the average response map difference when compared to the baseline BACF, thus providing empirical evidence of aberrance mitigation.

Theoretical and Practical Implications

The introduction of ARCF poses several implications for both theoretical explorations and practical applications:

1. Theoretical Advancement: ARCF extends the theoretical framework of DCF by integrating a mechanism that addresses aberrance directly within the learning process. This establishes a foundation for further improvements in DCF-based methodologies and applications where maintaining precise and reliable tracking is imperative, such as autonomous drones conducting complex tasks.
2. Practical Application: The practical significance of ARCF is underscored by its ability to handle dynamically variable environments inherent in UAV operations. The method offers potential enhancements in UAV systems for applications including surveillance, traffic monitoring, and search and rescue operations, where real-time object tracking accuracy is critical.

Future Developments

While ARCF currently employs hand-crafted features such as HOG and CN for feature extraction given its real-time application focus, future developments could involve integrating more sophisticated deep learning features to further augment its precision without compromising speed unduly. Additionally, the proposed aberrance repression framework shows promise for integration into other advanced trackers like ECO and SRDCF, indicating a potential universal application across various tracking systems.

This paper enriches the repertoire of DCF extensions by providing a viable solution to one of its critical flaws. It paves the way for ongoing enhancements in correlation filter-based tracking methodologies, contributing valuable insights into maintaining response map integrity amidst tracking challenges.

In conclusion, the ARCF method stands as a significant contribution to the continuous refinement of object tracking systems, particularly where performance in unpredictable and constrained environments such as those experienced by UAVs is prioritized.