Evidential Detection and Tracking Collaboration: New Problem, Benchmark and Algorithm for Robust Anti-UAV System

Problem formulation

• Prior anti-UAV work treats the task as tracking with UAV information provided, but this fails when UAVs appear randomly in complex scenes.The proposed task removes prior UAV information and requires continuous perception in open-world videos.Source: §I · p. 2

Limitations of existing benchmarks

• Existing datasets require a UAV in the first frame, use relatively large targets, and have limited quality and scale for real-world evaluation.These conditions do not represent UAVs appearing anywhere or occupying very few pixels.Source: §I · p. 2

AntiUAV600 benchmark

• AntiUAV600 provides 600 thermal infrared video sequences and over 720K densely annotated frames covering random appearance, fast motion, thermal crossovers, and tiny UAVs.Its sequences include practical cases where UAV targets do not appear at the beginning of videos.Source: §I · p. 2

Evaluation metric

• The benchmark introduces an evaluation metric combining average Intersection over Union with a penalty for evaluating localization and UAV absence prediction.The metric is designed for videos in which targets may appear or disappear.Source: §I · p. 2

Evidential detection-tracking collaboration

• The proposed framework uses evidential collaboration between global UAV detection and local UAV tracking to handle random appearance and absence.It begins with detection, switches to tracking after localization, and uses an evidential head to estimate tracking confidence for adaptive switching.Source: §I · pp. 2–3

Benchmark comparison

• Compared with existing UAV datasets, AntiUAV600 contains more complex scenarios, smaller average UAV scales, greater scale variation, and challenges that tracking methods struggle to handle.These comparisons are illustrated across Fig. 1(a-c).Source: §I · p. 3

Results

• Adaptive switching between detection and tracking achieves SOTA performance on AntiUAV600 and provides a strong baseline for future research.The paper reports extensive experiments and analyses supporting the superiority of the benchmark and framework.Source: §I · p. 3

Anti-UAV Datasets and Methodologies

• Prior datasets include visible-image detection and video-tracking benchmarks, while Anti-UAV provides 318 videos with over 297K annotated RGB-Thermal frames.These datasets support UAV perception research but commonly retain separate detection or tracking settings.Source: §II · p. 4
• Existing tracking methods combine global detection, local tracking, appearance, motion, background correlation, and dynamic small-target detection to handle challenging UAV scenes.The cited approaches target dynamic backgrounds, small UAV scales, and robust re-detection.Source: §II · p. 4

• Existing anti-UAV datasets and solutions typically address detection or tracking separately, although practical systems require adaptive detection and continuous target tracking.The paper therefore formulates a unified anti-UAV task and solution.Source: §II · p. 4

Evidential Deep Learning (EDL)

• Evidential Deep Learning models predictive uncertainty and can distinguish known from unknown data, enabling uncertainty-based confidence scoring.The paper uses this capability to assess whether a local tracker remains reliable.Source: §II · p. 4
• The proposed evidential head acts as a switcher that identifies tracking uncertainty and enables adaptive collaboration between detection and tracking.This design addresses the need to determine when local tracking is effective and when global detection should take over.Source: §II · p. 4

III. NEW ANTI-UAV BENCHMARK

• The new anti-UAV task continuously detects and tracks UAVs in real-time complex video streams without relying on a first-frame UAV template.Systems must determine UAV presence or absence and localize targets when present.Source: §III · p. 5

A. Dataset Details

• AntiUAV600 contains 600 sequences split into 300 training, 50 validation, and 250 test videos, with 337K, 56K, and 330K thermal infrared frames, respectively.Frames were captured at 640×512 resolution and 25 frames per second.Source: A. Dataset Details · p. 5
• AntiUAV600 is more challenging than existing datasets because it combines complex backgrounds, small UAV scales, and random UAV appearance or re-appearance.Its scenes include skies, city buildings, mountains, rivers, and forests, while targets may appear or disappear throughout a video.Source: A. Dataset Details · p. 5

B. Data Annotation

• The dataset provides dense bounding-box annotations for frames containing UAVs, labels frames without UAVs as “Not Exist,” and records seven challenging attributes.The attributes include Out-of-View, Occlusion, Fast Motion, Scale Variation, Infrared Crossover, Dynamic Background Clusters, and Target Scale.Source: B. Data Annotation · pp. 5–6
• AntiUAV600 has diverse UAV trajectories across the field of view and similar small-target scale distributions across its splits, creating significant perception challenges.Its average target size is smaller than those of other UAV datasets.Source: B. Data Annotation · p. 6

C. Evaluation Metric

• Acc evaluates anti-UAV performance using IoU-based localization accuracy together with penalties for visibility and localization failures.It averages performance over all frames while separately accounting for frames in which the UAV is visible.Source: C. Evaluation Metric · p. 6
• The metric penalizes missed UAV presence because failing to detect a visible target can impose high practical costs on anti-UAV systems.This explicitly distinguishes localization quality from failures to recognize that a target exists.Source: C. Evaluation Metric · p. 6

IV. EVIDENTIAL DETECTION AND TRACKING COLLABORATION FOR ANTI-UAV SOLUTION

• Evidential Detection and Tracking Collaboration (EDTC) uses a global detection branch for field-wide UAV detection and a local tracking branch for continuous localization.The two branches are adaptively switched for efficient and accurate perception in complex videos without prior UAV information.Source: §IV · p. 7

A. Global Detection

• EDTC detects UAVs throughout the video, switches to local tracking after localization, and returns to global detection when tracking is denied until the video ends.This cycle supports UAV reappearance without requiring a UAV in the first frame.Source: A. Global Detection · p. 7

B. Local Tracking

• The local tracking branch combines a three-stage relevance decoupling module, a corner head for bounding-box prediction, and an evidential head for tracking assessment.The corner head estimates target position and scale, using a training loss that combines IoU and L1 losses.Source: B. Local Tracking · pp. 7–8
• The relevance decoupling module separates template and search tokens, extracts local spatial features, and computes self- and cross-attention for feature matching.Template-independent search features support local UAV detection, while matched template–search features support tracking.Source: B. Local Tracking · p. 8
• The evidential head models target-versus-background tracking as a binary classification problem and predicts class probability together with uncertainty.Evidence is converted into Dirichlet parameters, class probabilities, and predictive uncertainty.Source: B. Local Tracking · pp. 8–9
• EDTC continues tracking when the target probability has uncertainty below θeh and switches to global detection otherwise.The evidential head compares search-region prediction information with the initial target before producing the evidence used for this decision.Source: B. Local Tracking · p. 9

V. EXPERIMENTS

• The detection branch is trained on AntiUAV600, while the tracking branch is trained on AntiUAV600 plus LaSOT [10], TrackingNet [20], GOT10K [13], and COCO [17].The branches are optimized separately.Source: A. Implementation Details · p. 9
• 35 fps is the reported inference speed of EDTC, with α=0.2 and β=0.3 in Eq. 1.Additional implementation details are provided in the supplementary material.Source: A. Implementation Details · p. 9

B. Comparative Experiments

• The sequential comparison combines a UAV-specific YOLOv5s detector with transformer-, Siamese-, and filter-based trackers, while EDTC and EDTC⋆ differ by whether UAV data train the tracking branch.The compared trackers include OSTrack [32], MixFormer, ToMP [18], Stark [31], KeepTrack [19], SiamCAR [11], and DiMP [3].Source: B. Comparative Experiments · p. 9
• EDTC’s evidential collaboration achieves significantly superior performance to the compared approaches on AntiUAV600.The comparison is reported in Table II.Source: B. Comparative Experiments · p. 10, Table II · p. 10
• 0.606 Acc is achieved by EDTC⋆ on DUT Anti-UAV, improving over EDTC’s 0.593 Acc when the additional AntiUAV600 training data are used.The result is reported for the thermal-video evaluation setting.Source: B. Comparative Experiments · p. 10, Table III · p. 10
• 0.634 Acc is achieved by EDTC⋆ on Anti-UAV after adding UAV data for training, compared with 0.617 for EDTC.EDTC outperforms the other compared trackers on this dataset.Source: B. Comparative Experiments · p. 10, Table III · p. 10
• EDTC improves UAV perception in videos with long target absences and reappearances, where existing tracking methods struggle.The evaluation uses 63 AntiUAV600 test videos whose targets are present initially and later disappear and reappear.Source: B. Comparative Experiments · p. 10, Table IV · p. 10

C. Ablation Studies

• Comparing Tracker2 with Tracker1, adding the AntiUAV600 dataset for training improves Acc in SC from 0.263 to 0.278 and in EC from 0.439 to 0.449.These results emphasize AntiUAV600’s advantageous contribution to UAV perception.Source: C. Ablation Studies · p. 11
• AntiUAV600 produces SC Acc values of 0.278 and 0.352 and EC Acc values of 0.449 and 0.486 in the reported tracker comparisons.These comparisons contrast Tracker1, Tracker2, and Tracker3 trained with general, Anti-UAV, and AntiUAV600 data.Source: C. Ablation Studies · p. 11
• The ablation isolates the global detection branch, tracking branch, and Evidential Head as key components of EDTC.BaseTracker denotes the tracking branch without the evidential head.Source: C. Ablation Studies · p. 11
• 0.486 Acc is achieved by integrating global detection and local tracking through the evidential head, improving over 0.392 for global detection alone.The simple detector-to-tracker switch instead drops performance from 0.392 to 0.352 because false positives can make the tracker follow distractors.Source: C. Ablation Studies · pp. 11–12
• The evidential head switches to global detection when a target prediction has uncertainty above θeh, treating such predictions as possible false positives.True-positive tracking results typically have low uncertainty.Source: C. Ablation Studies · p. 12
• 0.486 is the best reported performance at θeh=0.2, while using only evidential classification at θeh=0.05 still improves Acc from 0.352 to 0.468.Performance drops as θeh increases, supporting the usefulness of the predicted uncertainty.Source: C. Ablation Studies · p. 12

VI. CONCLUSION

• The new task requires systems to continuously detect UAV presence and accurately localize targets using only real-time captured videos.Source: §VI · p. 12
• AntiUAV600 and a novel performance measure provide a large-scale benchmark for the newly formulated anti-UAV problem.The measure accounts for both localization accuracy and the ability to predict UAV absence.Source: §VI · p. 12
• The proposed approach dynamically switches global detection and local tracking through an evidential learning module.Source: §VI · p. 12
• Comprehensive experiments demonstrate the effectiveness and advantage of the proposed solution and problem formulation for UAV perception.Source: §VI · p. 12

DATASET DETAILS

• AntiUAV600 covers diverse backgrounds and challenges including UAV absence and reappearance, fast motion, dynamic backgrounds, infrared crossover, and small scales.Scenes include cloudy skies, urban environments, mountains, rivers, and forests.Source: DATASET DETAILS · p. 15
• Because UAVs may not appear in the first frame, standard tracking alone is insufficient, motivating collaboration between detection and tracking.Source: DATASET DETAILS · p. 16
• Every frame is annotated with seven attributes: out-of-view, occlusion, fast motion, scale variation, infrared crossover, dynamic background clusters, and target scale.Source: DATASET DETAILS · p. 16
• Frequent absence, fast motion, thermal crossover, background clutter, and tiny size are the dataset’s main UAV-perception challenges.Source: DATASET DETAILS · p. 16

A. Implementation Details

• The framework combines a YOLOv5s global detector with a local tracking branch that uses template and search-region inputs to estimate target position.Source: A. Implementation Details · p. 16
• During inference, the detector processes the entire image, selects the highest-scoring UAV prediction, and supplies the detected target template and search region to tracking.Source: A. Implementation Details · p. 17

B. Qualitative Results.

• The proposed method achieves more accurate and steady UAV localization than global YOLOv5s detection and a simple detector-tracker combination.Source: B. Qualitative Results · p. 17
• Evidential detection-tracking collaboration improves UAV perception in challenging sequences involving target absence and reappearance, tiny scale, infrared crossover, and cluttered backgrounds.Source: B. Qualitative Results · p. 17