MUVAD: UAV Anomaly and Moving Object Datasets

• MUVAD is a dual-dataset resource that supports both video anomaly detection and moving object detection with UAV-captured footage.
• The datasets feature diverse real-world conditions, including continuous camera motion, complex backgrounds, and detailed frame-level or bounding box annotations.
• Benchmark evaluations using metrics such as AUC, precision, and recall highlight the datasets' value in advancing motion-robust computer vision methods.

The Moving UAV VAD Dataset (MUVAD) refers to two distinct publicly available datasets constructed for moving camera Video Anomaly Detection (VAD) and moving-object detection in dynamically captured Unmanned Aerial Vehicle (UAV) video. Both datasets are designed to address limitations of static-camera surveillance benchmarks by introducing real-world scenarios characterized by UAV-induced ego-motion, multi-source motion coupling, and complex backgrounds. MUVAD datasets serve as benchmarks for evaluating algorithms concerning anomaly detection and small-object detection in urban and natural environments, representing high-value resources for the development and validation of motion-robust computer vision methods (Liu et al., 16 Jan 2026, Delibasoglu, 2021).

1. Dataset Definitions and Scope

Two principal datasets known as MUVAD are referenced in current literature:

(A) MUVAD (Video Anomaly Detection) (Liu et al., 16 Jan 2026):

A large-scale, dynamic UAV video dataset created to evaluate VAD algorithms under realistic aerial surveillance conditions with continuously moving camera platforms, focusing on urban traffic anomaly detection.

(B) MUVAD (PESMOD; Moving Object Detection) (Delibasoglu, 2021):

A high-resolution UAV video dataset curated for benchmarking moving-object detection methods where scenarios present challenging small-target detection amidst free-form 6-DoF camera motion, varied environments, and no predefined train/test split.

Both datasets share the goal of advancing detection methodologies capable of handling UAV-induced background variability but differ in annotation granularity, event taxonomy, and intended evaluation protocols.

2. Data Acquisition Methodologies

(A) MUVAD (Video Anomaly Detection)

• Flight Patterns: Continuous UAV translation, rotation, and object-tracking maneuvers over urban arterials, intersections, hubs, and construction zones.
• Capture Protocol: Normal (training) sequences are obtained via routine UAV patrols; test (anomalous) sequences are collected both on-board and via curated YouTube videos, only retaining those with adequate quality, duration, and exclusive UAV capture.
• Resolution and Frame Rate: 852 × 480 pixels, 30 fps.
• Environmental Conditions: Diverse lighting (day/night) and weather (sunny/foggy) patterns are represented; all scenes feature substantial background dynamics resulting from UAV movement.

(B) MUVAD (PESMOD; Moving Object Detection)

• Source Footage: Eight video sequences, sourced from freely available drone videos on Pexels.
• Platform: Consumer-grade UAVs (models and flight altitudes unspecified) with standard RGB sensors.
• Resolution: 1920 × 1080 pixels (Full HD).
• Environments: Urban roads, trekking trails, snow fields, and rural landscapes under daytime lighting.
• Motion Dynamics: Both slow and rapid UAV-induced camera motion; some sequences include zoom, pitch, and roll changes.

3. Annotation Schemes and Quality Control

(A) MUVAD (Video Anomaly Detection)

• Labeling Protocol: Frame-level binary labels (normal = 0, anomaly = 1). Anomaly intervals correspond to the visibility of any of 12 predefined anomalous event types.
• Annotation Modality: No bounding box or pixel segmentation; only framewise annotation.
• Quality Assurance: Each test video sequence is independently labeled by multiple annotators, with disputed frames reviewed by domain experts before finalization.

(B) MUVAD (PESMOD; Moving Object Detection)

• Labeling Protocol: Axis-aligned bounding boxes for single-class "moving object" (including pedestrians and vehicles), with one bounding box per instance.
• Annotation Method: Fully manual; specifics on inter-annotator agreement or additional validation procedures are not detailed.
• Granularity: No segmentation masks provided; annotation formats likely text or CSV per frame.

4. Dataset Statistics and Splits

(A) MUVAD (Video Anomaly Detection)

Split	Videos	Frames	Anomaly Events	Anomaly Types	Background Type
Training	46	126,254	0	0	Dynamic
Testing	72	96,482	240	12	Dynamic
Total	118	222,736	240	12	Dynamic

• Anomaly Density: Less than 1% of total frames are labeled anomalous.

(B) MUVAD (PESMOD; Moving Object Detection)

Sequence Name	Frames	Moving-Object Boxes
Pexels-Elliot-road	664	3,416
Pexels-Miksanskiy	729	189
Pexels-Shuraev-trekking	400	800
Pexels-Welton	470	1,129
Pexels-Marian	622	2,791
Pexels-Grisha-snow	115	1,150
Pexels-Zaborski	582	3,290
Pexels-Wolfgang	525	1,069
Total	4,107	13,834

• Evaluation: No formal train/test split; all methods are benchmarked on all sequences.

5. Task Definitions and Evaluation Metrics

(A) MUVAD (Video Anomaly Detection)

• Anomaly Categories: 12 types, including illegal lane change, emergency lane violation, wrong-way driving, construction zone, vehicle breakdown, animal intrusion, vehicle skidding, vehicle collision, fire incident, roadside deviation, traffic congestion, and pedestrian intrusion.
• Protocol: One-class training—models are trained exclusively on normal examples and tested on both normal and anomalous frames.
• Metrics: Micro-AUC and Macro-AUC are recommended for performance reporting; usage guidelines emphasize adherence to this protocol.

(B) MUVAD (PESMOD; Moving Object Detection)

• Detection Criteria: A detected box $R_{\rm DET}$ is a true positive (TP) if $$\frac{\mathrm{area}(R_{\rm GT} \cap R_{\rm DET})}{\mathrm{area}(R_{\rm GT})} \geq 0.2$$.
• Evaluation Metrics:
  • Precision, recall, F1-score (standard definitions).
  • Overlap ratio $$O_r = \frac{\mathrm{area}(R_{\rm GT} \cap R_{\rm DET})}{\mathrm{area}(R_{\rm GT})}$$ for true positives.
• Baselines:
  • MCD (Yi et al., 2013): Grid-based KLT/RANSAC homography with Gaussian background model and neighbor mixing.
  • SCBU (Yun et al., 2017): Motion-compensation backbone with scene-conditional background updating.
  • Proposed method: Lightweight, flow-weighted differencing with adaptive thresholds and neighborhood differencing.

Comparison of methods (average across all 8 sequences):

Metric	MCD	SCBU	Proposed
$O_r$	0.7756	0.4949	0.6755
Precision	0.5428	0.4268	0.5452
Recall	0.5607	0.4416	0.5070
$F_1$	0.3828	0.4108	0.4302

6. Position Among Related UAV Datasets

(A) MUVAD (Video Anomaly Detection)

Dataset	Videos	Frames	Events	Anomaly Types	Background
CHUK Avenue	16/21	30,652	77	5	Static
ShanghaiTech	238/199	317,398	158	11	Static
Drone-Anomaly	37/22	87,488	26	10	35% Dynamic
UIT-ADrone	41/51	206,194	1,935	10	Static
MUVAD	46/72	222,736	240	12	Dynamic

• MUVAD offers a fully dynamic background, comparable in scale to the largest existing datasets, with broader anomaly coverage and more realistic moving-camera scenarios (Liu et al., 16 Jan 2026).

(B) MUVAD (PESMOD; Moving Object Detection)

• Resolution and Target Scale: Full HD, with small target size (tens of pixels) compared to datasets like VIVID and CDnet.
• Camera Motion: Supports free 6-DoF UAV maneuvers, beyond CDnet's Pan-Tilt-Zoom or VIVID’s largely planar motion.
• Advantages: Provides high-resolution, diverse real-world footage for evaluating detection under strong background variability.
• Limitations: Only bounding-box annotations, lack of night/IR imagery, no camera or flight metadata (Delibasoglu, 2021).

7. Access, Licensing, and Usage Recommendations

(A) MUVAD (Video Anomaly Detection)

• Repository: https://github.com/uavano/FTDMamba
• Citation: "FTDMamba: Frequency-Assisted Temporal Dilation Mamba for Unmanned Aerial Vehicle Video Anomaly Detection, C.-Z. Liu et al., IEEE Transactions on [Journal], 2025."
• Usage: Follow one-class training protocol (normal-only training), report Micro/Macro-AUC, adhere strictly to ethical usage—exclude privacy-infringing footage (Liu et al., 16 Jan 2026).

(B) MUVAD (PESMOD; Moving Object Detection)

• Shared under open terms as per (Delibasoglu, 2021); further details available in the associated publication.
• Intended as a motion-robust detection benchmark for academic research, emphasizing object detection amid UAV-induced ego-motion.

---

The MUVAD datasets collectively provide critical benchmarks for advancing UAV-based anomaly and object detection under unconstrained camera motion, enabling rigorous evaluation of algorithms in dynamic, real-world environments (Liu et al., 16 Jan 2026, Delibasoglu, 2021).