Bounding-Box Trajectories Matter for Video Anomaly Detection

Accepted to ECCV 2026

Inpyo Song and Jangwon Lee
Sungkyunkwan University
Paper (soon) arXiv Code (soon) Lab
TrajVAD contrasts conventional pose-based VAD with trajectory and pose VAD.
Conventional pose-based VAD discards trajectory signals after detection and tracking. TrajVAD treats bounding-box trajectories as the core anomaly signal across object classes.

Abstract

Video anomaly detection is critical for public safety and security, yet remains highly challenging despite extensive research due to large variations in appearance, viewpoint, and scene dynamics. Among existing approaches, human pose-based methods have emerged as a major line of research, showing strong performance since many anomalies in public datasets involve humans and pose representations are robust to appearance changes while providing compact motion descriptions. However, these methods often overlook bounding-box trajectories, although such information is inherently available in pose-based pipelines. In this paper, we explicitly leverage these trajectories as a primary anomaly cue. We present TrajVAD, a framework that models multi-class bounding-box trajectories using normalizing flows to learn normal kinematic patterns. Its trajectory-only variant (TrajVAD-T) eliminates pose estimation and surpasses all compared pose-based methods on ShanghaiTech in AP (87.7%), while achieving the best results on MSAD. An extended version (TrajVAD-P) incorporates pose information and further improves performance to 88.6% AUROC and 90.9% AP on ShanghaiTech, highlighting bounding-box trajectories as an effective yet underexplored modality for video anomaly detection.

Method

TrajVAD starts from standard detection and multi-class tracking. Each track is converted into a sequence of 27 trajectory-derived features, including box state, velocity, acceleration, direction, scale change, perspective-normalized motion, and detector confidence. A class embedding lets a shared normalizing flow learn category-specific normal motion patterns.

TrajVAD-T scores trajectories using negative log-likelihood under the learned normal distribution. TrajVAD-P adds a conditional pose flow only for person tracks with reliable keypoints, so unreliable skeletons do not dominate the anomaly score.

TrajVAD architecture with trajectory normalizing flow and optional pose branch.
Multi-class tracks are encoded as trajectory feature sequences and conditioned on class embeddings. The optional pose branch is gated by pose reliability.

Results

Method Venue SHT AUROC SHT AP UBn AUROC UBn AP MSAD-HR AUROC MSAD-HR AP
STG-NF ICCV'23 85.9 77.6 71.8 62.7 55.7 56.5
SeeKer ICCV'25 85.5 80.0 77.9 80.3 61.1 60.1
TrajVAD-T ECCV 2026 84.9 87.7 68.0 63.2 69.7 60.4
TrajVAD-P ECCV 2026 88.6 90.9 73.8 68.3 68.5 58.5

SHT: ShanghaiTech. UBn: UBnormal. MSAD-HR: human-related subset of MSAD.

Computational Efficiency

Method Venue Det+Track
(ms/frame)		 +Pose
(ms/frame)		 Inference
(ms/seg)		 Total
(ms/seg)		 SHT-HR AUROC SHT-HR AP
TrajREC WACV'24 105.8 31.0 4.5 141.3 77.9 -
STG-NF ICCV'23 105.8 31.0 8.8 145.6 87.4 81.4
MoCoDAD ICCV'23 105.8 31.0 2,857.0 2,993.8 77.6 -
GiCiSAD WACV'25 105.8 31.0 1,296.1 1,432.9 78.0 -
TrajVAD-T ECCV 2026 104.4 - 2.8 107.2 84.1 87.5
TrajVAD-P ECCV 2026 104.4 31.0 8.6 144.0 88.6 91.1

Preprocessing is measured per frame and inference per segment. Total follows stride-1 sliding-window evaluation, where one new segment is produced per frame.

Qualitative Results

ShanghaiTech example: non-person anomaly scored from bounding-box trajectory dynamics.
MSAD example: traffic accident detection benefits from multi-class trajectory coverage.
Qualitative comparison of TrajVAD-T, TrajVAD-P, and STG-NF.
TrajVAD detects anomalies from bounding-box kinematics when pose-only scoring has no signal for vehicles or degraded skeletons.

BibTeX

@article{song2026bounding,
  title={Bounding-Box Trajectories Matter for Video Anomaly Detection},
  author={Song, Inpyo and Lee, Jangwon},
  journal={arXiv preprint arXiv:2605.21957},
  year={2026}
}