YUTELA: A Scalable Edge Cloud AIoT System for Multi-Site Video Anomaly Detection with Scene Specific Adaptation
Authors: Shanle Yao, Babak Rahimi Ardabili, Amir Kimiyaie, Hamed Tabkhi
Organizations: The authors are with the Electrical and Computer Engineering Department, The University of North Carolina at Charlotte, Charlotte, NC, 28223 USA.
Building upon our previous Ancilia framework, this work presents YUTELA, an edge--server--cloud AIoT system for long-term, multi-site video anomaly detection. YUTELA replaces workstation-based on-site processing with NVIDIA Jetson Thor edge devices and redesigns the multi-camera pipeline through camera batching and redistribution, shared memory, reduced inter-stage data transfer, task integration, and TensorRT-accelerated inference. It maintains a separate anomaly model per camera, retrained on a centralized GPU server once enough recent pseudo-normal data are collected, with updated weights returned to the edge. With eight cameras, TensorRT increases throughput from 14.17 to 29.94 FPS under normal density, and the optimized edge node exceeds the throughput of the workstation-based Ancilia system in seven of nine workload settings, reaching up to 2.03 times its throughput at extreme density. This gain comes at the cost of higher latency under lighter loads with four or eight streams. YUTELA was then operated continuously for 170 hours at four public sites using 31 existing CCTV cameras. The deployment shows that runtime load is driven by scene activity and input frame rate rather than by camera count, and that retraining frequency varies by about three times across sites. The 101 automatically triggered training events occupied only about 6.6% of a single GPU, indicating that one global node can support adaptation for tens of sites. Mean end-to-end alert latency ranged from 6.8 to 15.4~s depending on site traffic and anomaly type. These results show that YUTELA enables practical multi-camera edge processing with low-cost, camera-specific adaptation across heterogeneous sites.
Figures & tables
Fig. 1: YUTELA system architecture and data flow. N distributed edge nodes perform real-time video analysis and interact with a global node for scene-specific anomaly adaptation and multi-site statistical analysis, including behavior anomaly information, while anomaly information is transferred to the cloud for alerting and user interaction. Multiple deployment sites may share the global and cloud infrastructure, although one group is shown for clarity. BBP , IDP , KP , AS , BBO , D , SSTD , WW , I , SA , A , UI , R , and FB denote pedestrian bounding boxes, identities, keypoints, anomaly scores, object-anomaly bounding boxes, batched data, scene-specific trainable data, updated weights, database information, statistical analysis, alerts, user-interface information, requests, and feedback, respectively.
Task
Method
Performance
Dataset
High-Accuracy Reference Algorithms
Object Detection
Co-DETR [ 42 ]
66.0 (AP)
COCO [ 43 ]
Tracking
HyperSSM [ 44 ]
80.0 (MOTA)
MOT20 [ 45 ]
Pose Estimation
ViTPose [ 46 ]
81.1 (AP)
COCO [ 43 ]
Anomaly Detection
SPARTA [ 30 ]
64.9 (AUC-ROC)
HuVAD [ 41 ]
Ancilia Algorithmic Core [ 4 ]
TABLE I: Comparison of high-accuracy reference models and the algorithmic cores of YUTELA.
Fig. 2: YUTELA edge processing and scene-specific adaptation architecture. Batched multi-camera streams share edge memory and inference resources, while anomaly features are collected at the global node for scene-specific model retraining.
Fig. 3: Representative camera views from the four real-world YUTELA deployment sites, including a university campus, local market, downtown restaurant, and art center. Two camera views are shown for each site to illustrate differences in scene layout, viewpoint, and deployment environment.
Fig. 4: Edge-node throughput (top) and end-to-end pipeline latency (bottom) as the number of concurrent camera streams increases, under normal (70 detections/s), heavy (216 detections/s), and extreme (744 detections/s) crowd density. The Ancilia workstation (Threadripper Pro 3975WX, three A6000 GPUs) is compared with YUTELA on a single Jetson Thor, with and without TensorRT.
Site
Deployment
Runtime Performance
Triggered Continual Training
Cams.
Env.
Input FPS
Peak Flow
Min Proc. FPS /Cam.
Max Lat. (s) /Cam.
Total Triggers
Triggers /Cam.
Inter-Trigger Interval (h)
Train Time /Event (min)
Site A
8
Urban Campus
30
29
10.96
16.71
9
0–3
55.0 ± 20.4
4.2 ± 2.4
Site B
7
Dining Complex
30
63
2.98
37.38
59
6–10
19.4 ± 4.6
7.7 ± 4.4
Site C
8
Public Marketplace
15
77
2.78
28.03
20
1-4
42.2 ± 6.3
6.2 ± 3.4
Site D
8
Center
20
30
1.92
15.98
13
0-5
41.0 ± 12.8
4.6 ± 3.1
TABLE II: Multi-site deployment, runtime performance, and triggered continual-training statistics.
Fig. 5: Runtime behavior of YUTELA during the continuous 170-hour deployments at the four sites. In each panel, the upper plot shows processing throughput (FPS) and pipeline latency, and the lower plot shows the number of people detected across the site’s cameras.
System
Location
Traffic Flow
Anomaly
Mean ± Std. (s)
Range (s)
Ancilia [ 8 ]
Location B
0–10
Object
8.92±0.71
7.46–10.61
Behavior
11.62±0.62
10.75–13.30
YUTELA
Site A
0–10
Object
6.82±0.57
5.70–7.69
Behavior
8.71±0.63
7.87–10.27
Site B
20–30
Object
11.33±0.69
9.98–12.39
Behavior
15.41±0.74
13.96–16.54
TABLE III: Physical-Cyber-Physcial alert-trigger latency for object and behavior anomalies with eight concurrent surveillance cameras under different traffic flows. The Ancilia baseline is reported from [ 8 ] .
Video anomaly detection in surveillance settings must balance detection accuracy against real-time throughput, a tension that existing methods address either through stronger feature extractors or more efficient architectures, but rarely both. We present VigilFormer, a unified framework that combines deformable spatio-temporal attention with causal temporal modeling to detect anomalies in untrimmed surveillance video. The proposed Deformable Spatio-Temporal Encoder (DSTE) attends to a sparse set of informative locations across frames, avoiding the quadratic cost of dense attention while retaining the ability to capture irregular motion patterns. A Causal Anomaly Classifier (CAC) applies dilated causal convolutions over snippet-level features and optimizes a contrastive multiple-instance learning objective that separates anomalous and normal representations without frame-level labels. To meet deployment constraints, an Adaptive Confidence Scheduler (ACS) dynamically skips low-information frames at inference time, reducing redundant computation in static scenes. Evaluated on UCF-Crime, ShanghaiTech, and CUHK Avenue, VigilFormer achieves AUC scores of 87.83%, 97.21%, and 89.74% respectively, at 41.5 FPS on a single GPU, outperforming recent weakly-supervised methods in both accuracy and speed.
Xinze Zhang
University of Southern California, Los Angeles, CA 90007, USA
Industrial inspection needs zero-shot anomaly detection (ZSAD) that remains useful under edge deployment constraints. Recent methods often rely on ViT-L foundation backbones (~300M parameters), which exceed the memory and operator budget of typical embedded hardware. We study this regime through EdgeZSAD, a compact reference system built around a TinyViT-21M-512 backbone, an asymmetric global-local readout (EdgeGLR), and a reproducible source-side training recipe (Real-IAD-DR). We train a single checkpoint in a source-trained, target-unseen protocol and evaluate it across six industrial benchmarks. Across three independent runs, the resulting model reaches an average image AUROC of 91.6 on MVTec-AD and 88.2 on VisA, while remaining directly deployable on Jetson Orin Nano Super (TensorRT FP16) and RB5 Gen2 (QNN GPU FP16). Across the six device-rescored benchmarks, image-AUROC drift stays below 0.2 points, indicating that the exported graph preserves host-side ranking behavior in the evaluated deployment setting.
Continuous AI inference on resource-constrained edge hardware introduces deployment effects that are largely invisible to conventional benchmark evaluation, including temporal instability in streaming video, thermal throttling under sustained load, and workload-dependent performance variability. We present Edge-TSR, a deployment-oriented continuous edge inference system for sustained roadside perception on the NVIDIA Jetson Orin Nano. Edge-TSR integrates detection, tracking, fine-grained classification, and a lightweight track-aware temporal stabilization mechanism that improves streaming inference consistency with negligible computational overhead. Our central finding is that benchmark-centric evaluation systematically overstates deployed edge inference performance. Across three state-of-the-art baselines, we observe consistent 20-30% relative degradation when transitioning from static-image evaluation to real-world streaming deployment. Edge-TSR addresses this gap through temporal inference stabilization, recovering up to 10.16% classification accuracy over per-frame inference baselines while maintaining sustained real-time performance under continuous operation. We evaluate the complete system under diverse real-world deployment conditions, jointly characterizing inference quality, latency, throughput, and thermal behavior during long-duration operation. A 55-minute vehicular deployment over a 26 km route demonstrates sustained operation at 16.18 FPS within safe thermal limits on a single embedded device without cloud offload. Our findings show that deployment-aware evaluation and temporal inference stabilization are necessary components of continuously operating edge AI systems intended for real-world sensing deployments. We release a sample annotated streaming video evaluation dataset and full system implementation to support reproducible deployment-centric evaluation.
Aditya Mishra, Haroon Lone
Indian Institute of Science Education and Research Bhopal, India