Source-linked AI summary
Learning Aberrance Repressed Correlation Filters for Real-Time UAV Tracking
Ziyuan Huang, Changhong Fu, Yiming Li, Fuling Lin, Peng Lu
TL;DR
The paper addresses DCF boundary effects and aberrances caused by excessive background information and appearance changes in UAV tracking. It proposes ARCF, which enlarges the search region while restricting response-map variations, achieving state-of-the-art performance with sufficient real-time speed.
Problem
UAV tracking requires robustness to fast motion, limited search regions, background clutter, occlusion, and appearance changes that can produce aberrances and tracking failures.
Method
ARCF combines a cropping matrix for search-region enlargement with a regularization term that restricts response-map changes during correlation-filter training.
Results
Across three UAV benchmarks, ARCF achieved state-of-the-art precision and success rate, substantially improved over BACF, and operated at sufficient speed for real-time tracking.
Takeaways & Limitations
Restricting response-map variations suppresses aberrances caused by background noise and tracked-object appearance changes, improving the robustness of DCF-based UAV tracking.
Abstract
from arXiv · showhide
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.
1. Introduction
UAV tracking is challenged by fast motion, occlusion, deformation, illumination variation, boundary effects, and aberrant response maps. ARCF addresses these issues by enlarging search regions, repressing response-map changes, and maintaining efficient tracking.
- Motivation: UAV tracking remains demanding because fast motion, occlusion, deformation, and illumination variation complicate robust and accurate object localization.
- Limitations of DCF: DCF gains computational efficiency through frequency-domain calculations but introduces artificial samples that cause damaging boundary effects.
- Aberrances: Traditional DCF generates response maps for localization but does not exploit their aberrance information, so tracking can be lost when aberrances occur.
- Contribution: ARCF combines a cropping matrix for search-region enlargement with a regularization term that suppresses abrupt response-map changes.
- Evaluation: ARCF was evaluated on 243 challenging UAV image sequences against 20 state-of-the-art hand-crafted and deep trackers.
- Conclusion: The authors report that aberrance repression improves DCF robustness and performance in UAV tracking while retaining sufficient speed for real-time use.
2. Related work
Prior DCF research addressed boundary effects through kernels, scale estimation, spatial regularization, and background information, while response-map quality received less attention. ARCF incorporates response-map restriction into a workflow using object and background samples.
- DCF developments: DCF research progressed from MOSSE filters and kernelization to scale estimation, context, background information, and hand-crafted features.
- Boundary-effect mitigation: SRDCF enlarges search regions by penalizing background during filter training but has high computational costs.
- Response-map methods: LMCF and ACFN use response-map quality for confidence-based updates or tracker selection, but both methods have limitations described in the related-work discussion.
- ARCF context: ARCF’s workflow learns from positive object samples and negative background samples while integrating response-map restriction into filter learning.
3. Background-aware correlation filter
BACF enlarges the search region by extracting real background patches through a cropping matrix, improving handling of relative motion. Excessive background information, however, increases clutter and aberrant detections.
- BACF formulation: BACF minimizes an objective over vectorized multichannel samples and an ideal response while learning a correlation filter for each channel.
- BACF formulation: The cropping matrix B ∈ R^M×N selects central M elements from each input channel, while w_d ∈ R^M is the filter learned for channel d.
- Search-region enlargement: Because M << N, BACF uses real background information rather than only periodically shifted patches during correlation-filter training.
- Benefit: The expanded search region helps track objects moving relatively quickly against the camera or UAV.
- Limitation: Excessive background information introduces clutter and makes similar contextual objects more likely to be detected as the target, producing aberrances.
4. Aberrance repressed correlation filter
ARCF suppresses abrupt response-map changes while preserving enlarged search regions, combining aberrance repression with efficient frequency-domain optimization.
- Aberrance identification: Euclidean response-map differences identify aberrances because sudden appearance changes reduce similarity and increase the difference measure.The previous response map is compared with the current map after shifting peaks to coincide.
- Aberrance repression: The third loss term restricts response-map changes between consecutive frames through an aberrance penalty γ during correlation-filter training.The restriction is transformed into the frequency domain so it can be optimized during training.
- Overall objective: ARCF retains BACF’s cropping matrix for search-region enlargement and adds regularization to counteract aberrances caused by background information.The cropping matrix preserves a sufficient search region, while the regularization term addresses the resulting aberrances.
- Frequency-domain formulation: The objective is transformed into the frequency domain because it remains a convolution operation, preserving sufficient computing efficiency.The transformed formulation introduces Fourier-domain variables, including the DFT-transformed signal and parameter ˆgk.
- Optimization through ADMM: ADMM solves the convex augmented-Lagrangian objective through two subproblems with closed-form solutions, while exploiting input sparsity to accelerate repeated calculations.The Sherman-Morrison formula further optimizes the inverse operation in the ˆgk subproblem.
- Appearance-model update: The appearance model is updated frame by frame using a learning rate η, with k and k −1 denoting consecutive frames.The previous-frame response map is treated as constant during the relevant frequency-domain calculation.
5. Experiments
ARCF is evaluated across three UAV benchmarks using hand-crafted and deep-feature trackers, with results showing strong performance, occlusion robustness, and real-time suitability.
- Experimental setup: 243 sequences comprising over 90,000 frames from UAV123@10fps, DTB70, and UAVDT are used to compare ARCF with 20 state-of-the-art trackers.Experiments include both hand-crafted and deep-based trackers.
- Overall performance: ARCF-HC outperformed all other hand-crafted feature-based trackers on UAV123@10fps, DTB70, and UAVDT.Figure 3 reports precision and AUC through precision and success plots.
- Overall performance: 0.666 precision and 0.472 AUC were achieved by ARCF-HC on UAV123@10fps, exceeding ECO-HC and STRCF in both metrics.The reported margins were 0.6% and 3.9% in precision, and 0.1% and 1.6% in AUC, over the second- and third-best trackers.
- Overall performance: 7.98% precision and 5.32% AUC improvements over BACF were reported for ARCF-HC, while ARCF-H improved by 2.77% and 0.69%.The experiments also report speed adequate for real-time UAV tracking applications.
- Attribute-based evaluation: 8.1%, 9.8%, and 5.2% AUC gains over BACF were obtained by ARCF-HC in occlusion cases on UAV123@10fps, DTB70, and UAVDT.The paper attributes this improvement to restricting response-map variations and limiting learning from irrelevant occluding objects.
- Qualitative and deep-based comparisons: ARCF remained competent under partial and full occlusions, while comparisons with deep-based trackers showed favorable precision and success-rate performance.Qualitative examples cover UAV123, DTB70, and UAVDT; Figure 7 reports quantitative comparison on UAVDT.
- Aberrance repression: 20%, 24%, and 15% lower average response-map differences were reported for ARCF-H than BACF on UAV123@10fps, DTB70, and UAVDT.The paper notes that ARCF suppresses most undesired fluctuations, although aberrances can still occur during out-of-view and full-occlusion cases.
6. Conclusion and future work
ARCF suppresses aberrances in UAV tracking while retaining real-time speed and improving precision and success rate over BACF. Its lightweight HOG and CN features support efficiency, while convolutional features are suggested when real-time demands are lower.
- Conclusion and future work: ARCF suppresses aberrances caused by BACF background noise and tracked-object appearance changes through response-map variation regularization.The regularization term restricts response-map variations during tracking.
- Conclusion and future work: ARCF achieves state-of-the-art precision and success rate on three UAV tracking benchmarks, with a substantial improvement over BACF.The evaluation covers three prevalent UAV benchmarks.
- Conclusion and future work: ARCF raises DCF tracker performance without sacrificing much speed, making its speed sufficient for real-time UAV tracking.The tracker uses HOG and CN features to preserve computing efficiency.
- Conclusion and future work: When real-time demands are lower, ARCF can incorporate more comprehensive convolutional features to improve precision and success rate.This is presented as a future extension of the framework.