Pose-based Video Anomaly Detection (VAD) has gained significant attention for its privacy-preserving nature and robustness to environmental variations. However, traditional frame-level evaluations treat video as a collection of isolated frames, fundamentally misaligned with how anomalies manifest and are acted upon in the real world. In operational surveillance systems, what matters is not the flagging of individual frames, but the reliable detection, localization, and reporting of a coherent anomalous event, a contiguous temporal episode with an identifiable onset and duration. Frame-level metrics are blind to this distinction, and as a result, they systematically overestimate model performance for any deployment that requires actionable, event-level alerts. In this work, we propose a shift toward an event-centric perspective in VAD. We first audit widely used VAD benchmarks, including SHT[19], CHAD[6], NWPUC[4], and HuVAD[25], to characterize their event structure. We then introduce two strategies for temporal event localization: a score-refinement pipeline with hierarchical Gaussian smoothing and adaptive binarization, and an end-to-end Dual-Branch Model that directly generates event-level detections. Finally, we establish the first event-based evaluation standard for VAD by adapting Temporal Action Localization metrics, including tIoU-based event matching and multi-threshold F1 evaluation. Our results quantify a substantial performance gap: while all SoTA models achieve frame-level AUC-ROC exceeding 52% on the NWPUC[4], their event-level localization precision falls below 10% even at a minimal tIoU=0.2, with an average event-level F1 of only 0.11 across all thresholds. The code base for this work is available at https://github.com/TeCSAR-UNCC/EventCentric-VAD.
Figures & tables
Aspects
Frame-Level
Event-Level
Unit of analysis
Individual frames
Temporal events
Consideration of temporal continuity
No
Yes
Boundary localization
Cannot measure
Measured via tIoU
Operational trust
Low
High
Sensitivity to class imbalance
High
Lower
Suitable for real-world deployment
Limited
Yes
Table 1: Comparison between frame-level and event-level evaluation paradigms for video anomaly detection (VAD).
Granularity
Characteristic
SHT [ 19 ]
CHAD [ 6 ]
HuVAD [ 25 ]
NWPUC [ 4 ]
Frame
Normal Frames
24,077
67,303
694,415
318,793
Anomalous Frames
16,714
59,172
225,075
65,266
Event
Anomalous Events
121
190
1,691
137
Avg. Duration (f)
138.13
311.43
133.10
476.39
Table 2: Frame-level and event-level statistics of VAD benchmarks.
Figure 1: The proposed three-stage Frame-to-Event Transformation framework. Raw anomaly scores undergo hierarchical Gaussian smoothing to surpass high-frequency noise. Adaptive thresholds ( τEER and τHprs ) are then applied to the smoothed signal to generate a binary output. Finally, a temporal refinement and short-event filter resolve fragmented detections to produce semantically coherent anomalous events (red boxes) aligned with human motion dynamics.
Figure 2: branch event-level anomaly detection framework. Given an input pose sequence, the model processes the data through two parallel branches: a Short Window Branch (SWB) with temporal length i and a Long Window Branch (LWB) with temporal length 3i . Both branches share the same transformer-based reconstruction backbone [ 24 ] , which jointly models absolute pose and relative pose through an encoder-decoder architecture. During inference, each branch produces frame-wise reconstruction errors. The center portion of the LWB error sequence is temporally aligned with the SWB target window, and the aligned frame-wise scores are fused to form a context-regularized anomaly response. The fused scores are then temporally pooled over the target window to produce a single event-level anomaly score.
STG-NF [ 15 ]
SPARTA [ 24 ]
TS-GAD [ 22 ]
Metric
SHT [ 19 ]
CHAD [ 6 ]
NWPUC [ 4 ]
HuVAD [ 25 ]
SHT [ 19 ]
CHAD [ 6 ]
NWPUC [ 4 ]
HuVAD [ 25 ]
SHT [ 19 ]
CHAD [ 6 ]
NWPUC [ 4 ]
HuVAD [ 25 ]
AUC-ROC
0.866
0.570
0.617
0.520
0.861
0.569
0.635
0.649
0.807
0.553
0.618
0.623
AUC-PR
0.838
0.569
0.276
0.252
0.831
0.517
0.258
0.327
0.740
0.499
0.300
0.305
EER
0.218
0.458
0.408
0.478
0.225
0.444
0.408
0.409
0.256
0.461
0.418
0.417
F1@τEER
0.746
0.52
0.330
0.347
0.738
0.538
0.329
0.413
0.703
0.521
0.322
0.406
F1@τHprs
0.744
0.518
0.323
0.377
0.729
0.536
0.332
0.405
0.692
0.521
0.322
0.421
Table 3: Frame-level evaluation performance of pose-based VAD models, STG-NF [ 15 ] , SPARTA [ 24 ] , TS-GAD [ 22 ] , across VAD benchmark datasets, SHT [ 19 ] , CHAD [ 6 ] , NWPUC [ 4 ] , HuVAD [ 25 ] . Thresholds τEER , and τHprs are extracted per model-dataset pair to support subsequent event-level analysis.
tIOU=0.5
tIOU=0.4
tIOU=0.3
tIOU=0.2
Data
Model
Prec.
Rec.
F1
Prec.
Rec.
F1
Prec.
Rec.
F1
Prec.
Rec.
F1
Average F1
SHT [ 19 ]
Baselines
STG-NF [ 15 ]
50.59%
71.07%
0.591
60.00%
84.30%
0.700
61.76%
86.78%
0.721
64.71%
90.91%
0.756
0.692
TS-GAD [ 22 ]
36.79%
58.68%
0.452
45.08%
71.90%
0.554
54.92%
87.60%
0.675
57.51%
91.74%
0.707
0.597
SPARTA [ 24 ]
44.62%
68.60%
0.540
52.15%
80.17%
0.631
56.45%
86.78%
0.684
0.596%
91.74%
0.723
0.644
Baselines + Proposed Post-Processing
Table 4: Event-level performance of pose-based VAD models, including STG-NF [ 15 ] , SPARTA [ 24 ] , TS-GAD [ 22 ] , and our proposed framework, across the SHT [ 19 ] , CHAD [ 6 ] , NWPUC [ 4 ] , and HuVAD [ 25 ] datasets, using τEER . Results are compared across three configurations: Baselines (frame grouping), Baseline+Post-processing (using our refinement pipeline), and our natively event-aware model.
tIOU=0.5
tIOU=0.4
tIOU=0.3
tIOU=0.2
Data
Model
Precision
Rec.
F1
Precision
Rec.
F1
Precision
Rec.
F1
Precision
Rec.
F1
Average F1
SHT [ 19 ]
Baselines
STG-NF [ 15 ]
51.79%
71.90%
0.602
57.14%
79.34%
0.664
58.93%
81.82%
0.685
63.69%
88.43%
0.740
0.673
TS-GAD [ 22 ]
41.38%
59.50%
0.488
50.00%
71.90%
0.589
57.47%
82.64%
0.678
62.07%
89.26%
0.732
0.622
SPARTA [ 24 ]
45.83%
63.64%
0.532
54.76%
76.03%
0.636
57.74%
80.17%
0.671
64.29%
89.26%
0.747
0.647
Baselines + Proposed Post-Processing
Table 5: Event-level performance of pose-based VAD models, including STG-NF [ 15 ] , SPARTA [ 24 ] , TS-GAD [ 22 ] , and our proposed framework, across the SHT [ 19 ] , CHAD [ 6 ] , NWPUC [ 4 ] , and HuVAD [ 25 ] datasets, using (τHprs) . Results are compared across three configurations: Baselines (frame grouping), Baseline+Post-processing (using our refinement pipeline), and our natively event-aware model.
Event-based vision, characterized by low redundancy, focus on dynamic motion, and inherent privacy-preserving properties, naturally fits the demands of video anomaly detection (VAD). However, the absence of dedicated event-stream anomaly detection datasets and effective modeling strategies has significantly hindered progress in this field. In this work, we take the first major step toward establishing event-based VAD as a unified research direction. We first construct multiple event-stream based benchmarks for video anomaly detection, featuring synchronized event and RGB recordings. Leveraging the unique properties of events, we then propose an EVent-centric spatiotemporal Video Anomaly Detection framework, namely EWAD, with three key innovations: an event density aware dynamic sampling strategy to select temporally informative segments; a density-modulated temporal modeling approach that captures contextual relations from sparse event streams; and an RGB-to-event knowledge distillation mechanism to enhance event-based representations under weak supervision. Extensive experiments on three benchmarks demonstrate that our EWAD achieves significant improvements over existing approaches, highlighting the potential and effectiveness of event-driven modeling for video anomaly detection. The benchmark datasets will be made publicly available.
Peng Wu, Yuting Yan, Guansong Pang +4
School of Computer Science, Northwestern Polytechnical University, China · School of Computing and Information Systems, Singapore Management University, Singapore · School of Artificial Intelligence, Xidian University, China
Video anomaly detection (VAD) aims to localize anomalous events by identifying when they occur. A common evaluation pools frame-level anomaly scores across all test videos and measures whether anomalous frames receive higher scores than normal frames (Micro-AUROC). This pooling creates comparisons both within the same video and across different videos. However, the pooled evaluation protocol commonly used in current VAD benchmarks makes Micro-AUROC heavily dependent on cross-video comparisons, which do not directly test whether anomalous frames rank above normal frames within the same video. To quantify this imbalance, we decompose Micro-AUROC into within-video and cross-video comparisons. Across four widely used weakly supervised VAD benchmarks, within-video pairs account for only 0.071-0.388% of all anomalous-normal frame pairs. This imbalance can make Micro-AUROC a poor indicator of within-video ranking. Video-level classifiers that assign a single score to each video achieve 81.40-97.18% Micro-AUROC while obtaining 50% Within-AUROC. For existing weakly supervised VAD detectors, replacing all frame scores in each video with their mean sets Within-AUROC to 50%. Across 120 runs, this replacement retains a median 98.8% of the original Micro-AUROC margin above chance. Because the decomposition depends on the evaluation protocol rather than the training supervision, the same distinction applies whenever VAD frame scores are pooled across videos. High Micro-AUROC alone therefore does not establish temporal localization. Our analysis reveals a limitation of pooled VAD evaluation and motivates reporting within-video ranking alongside Micro-AUROC.
Video Anomaly Detection (VAD) is inherently challenging due to the scarcity of anomalies and the large visual variability in surveillance footage, including changes in lighting, viewpoint, and human appearance. To mitigate visual noise and address privacy concerns, recent work has shifted to pose-based VAD, which focuses on motion dynamics rather than raw video data. However, existing pose-based approaches model human behavior in continuous latent spaces, limiting their ability to learn compact motion patterns necessary for robust behavior analysis. We address this by proposing Vector-Quantized Video Anomaly Detection (VQ-VAD), a novel human-centric anomaly detection framework that learns discrete motion representations. VQ-VAD adapts Vector-Quantized GAN (VQ-GAN), originally developed for image generation, to operate on keypoint sequences and construct a motion codebook of normal behavior. Trained exclusively on normal motion sequences, VQ-VAD detects anomalies by identifying high reconstruction errors when an observed motion sequence cannot be mapped to the learned codebook. We conduct extensive experiments across three complementary evaluation settings, including in-domain, cross-domain, and cross-dataset generalization, on four anomaly detection benchmarks. VQ-VAD achieves strong in-domain accuracy (81.83% on HR-SHT [15]), effective cross-domain transfer from CMU Panoptic [14] (76.69% on HR-SHT [15] without retraining), and competitive cross-dataset robustness. The code base for this work is available at https://github.com/TeCSAR-UNCC/VQ-VAD.
Narges Rashvand, Ghazal Alinezhad Noghre, Shanle Yao +2
Electrical and Computer Engineering Department University of North Carolina at Charlotte Charlotte, NC, USA