Machine Learning Video Anomaly Detection: 2026 Playbook for Security Teams

Why Fora Soft wrote this playbook

Fora Soft has built video and multimedia products since 2005. Over those 21 years we’ve shipped 250+ projects, and video surveillance, AI-assisted video analytics and real-time streaming are the spine of our portfolio. Our engineers wire up detection pipelines on top of WebRTC, RTMP/HLS, RTSP and NVIDIA DeepStream, run them on Jetson edge boards and AWS/Hetzner clusters, and keep a 100% on-time delivery record across those projects.

The clearest case study is V.A.L.T, our video surveillance and observation platform. V.A.L.T is deployed in regulated environments (law enforcement, clinical settings, a courtroom project in Kazakhstan) where false alerts are unacceptable and audit trails are mandatory. That operational reality—not a Kaggle leaderboard—shapes the recommendations in this guide.

The rest of this playbook answers the questions buyers actually ask us on intro calls: which model families work, what good architecture looks like, what the real costs are, where GDPR and the EU AI Act stop you, and when it’s cheaper to buy an off-the-shelf VMS instead of building one. Use it to scope your own project—or send us the spec and we’ll scope it with you.

What counts as a video anomaly in a security context

Before picking a model or a vendor, nail the anomaly taxonomy. Most “AI surveillance” failures happen because the buyer never separated “anomalous pixels” from “anomalous events” from “policy violations”. These are three different problems, with three different cost curves.

1. Object-level anomalies. Something that shouldn’t be there: a person in a restricted zone, a vehicle on a pedestrian path, an abandoned bag. Solvable with mature object-detection models (YOLOv8, RT-DETR) plus geofence rules. Commodity work.

2. Motion & behavior anomalies. Running in a no-running zone, loitering, a fall, a fight, a sudden crowd dispersion. These need spatio-temporal models (I3D, TimeSformer, Video Swin) and 3–10 seconds of context. Solvable but not trivial.

3. Contextual & policy anomalies. “This door should never open after 9 PM.” “This operator never logs in from outside the country.” These combine video events with access-control, schedule and identity data. The ML part is the easy bit; the integration with your existing systems is where 60% of the timeline goes.

Market snapshot: why every VMS vendor now ships an AI SKU

The AI video surveillance market sat at roughly $5.0–6.5 B in 2024 and is projected to reach $12–28 B by 2030, growing at a 20–31% CAGR depending on which analyst you believe (MarketsAndMarkets, Grand View Research). The broader video analytics segment—which includes retail, traffic and industrial use cases—hits about $37.8 B by 2030 at a 19.5% CAGR.

Three forces are driving that curve. First, edge silicon got cheap: a Jetson Orin Nano at $249 now runs models that needed a $5k workstation in 2020. Second, pre-trained backbones (I3D, Video Swin, SAM) collapsed the data-collection cost for buyers who don’t have a million labeled frames lying around. Third, vendors have proven they can cut operator workload—Motorola’s Avigilon reports ~90% reduction in false alarms with its self-learning Unusual Motion Detection.

The practical read: if you’re commissioning this product in 2026, you’re not on the bleeding edge, you’re adopting a tech stack that’s been production-hardened for 3–5 years. That’s why estimates are falling, not rising.

Model families that actually ship to production

Five architectures cover 95% of real deployments. Pick by how much labeled data you have, how tight your latency budget is, and whether your anomalies are spatial, temporal or both.

CNN backbones (ResNet, DenseNet, EfficientNet)

Why pick it. Fastest inference (2–5 TFLOP/frame), tiny memory footprint, every edge runtime supports it. Perfect for per-frame object or scene classification.

Limits. No temporal reasoning. A CNN cannot tell “a person walked into a bank” from “a person ran into a bank”. High false positives on static scenes with lighting flicker.

3D-CNN / I3D

Why pick it. Learns spatial + temporal features jointly. On UCF-Crime, I3D-based ensembles still sit at ~84.6% frame-level AUC—competitive with 2024 Transformer baselines. Mature tooling in NVIDIA TAO and DeepStream.

Limits. Memory-heavy: expect 6–8 GB VRAM per stream at 224×224, 32-frame clips. Slower than 2D alternatives on edge devices without TensorRT optimization.

Video Transformers (TimeSformer, Video Swin)

Why pick it. State of the art. Swin-3DART hit 0.861 ROC AUC on ShanghaiTech; SwinAnomaly (conditional GAN + Video Swin) holds near-SOTA while running real-time on Orin-class hardware. Long-range temporal context without LSTM’s vanishing-gradient pain.

Limits. Hungry for training data and compute; not the right pick if you have fewer than 10k labeled clips. Deployment requires careful attention to quantization to fit on edge devices.

Autoencoder & VAE reconstruction

Why pick it. Unsupervised. Train on hours of “normal” footage, flag frames the model can’t reconstruct. No labels needed. Ideal when your anomaly class is “anything weird” rather than a specific event.

Limits. Assumes reconstruction error correlates with abnormality—often false. Prone to high false positives on lighting changes, camera shake, new objects that are normal-but-rare.

Self-supervised & contrastive (MoCo, SimCLR, VideoMAE)

Why pick it. Few-shot friendly. Pre-train on large unlabeled video, fine-tune on a few hundred labeled anomaly clips. Cuts labeling budget by 10× in our experience.

Limits. Pre-training is expensive. Only worth it if you can amortize that cost across multiple deployments or sites.

Comparison matrix: which model for which job

Reference architecture: edge, cloud, or hybrid

The architecture decision dominates cost and latency far more than the model choice. Three patterns exist; hybrid wins for almost every serious deployment.

Edge-only

Detection runs on the camera or a local Jetson/Hailo module. Latency 10–100 ms. Bandwidth to headquarters is just alerts plus short event clips. Works offline. Best for fewer than 8 cameras per site or strict privacy regimes where raw video must stay on-premise.

Cloud-only

All streams go to AWS/GCP for inference. Latency 200–2000 ms. Bandwidth: ~5–50 Mbps per stream. Egress costs pile up fast; a 500-camera site at 4 Mbps average generates ~21 TB/day. Only justifiable for forensic search, low-rate polling, or when the customer has a dedicated backbone.

Hybrid (edge detection + cloud correlation)

Edge models filter the 99%+ of frames that are normal; only events and low-bitrate metadata travel to the cloud, where you do multi-camera correlation, long-term pattern mining, and dashboard/UX. Latency 50–300 ms. Bandwidth ~10% of raw streams. This is the default for everything Fora Soft ships today.

Edge hardware: what to put in the BOM

Your edge chip decides how many streams per box you can analyze and what models fit. In 2026 the practical shortlist is short.

For most of our clients, Jetson Orin Nano plus a local PoE switch covers 2–4 streams per appliance and clears all standard anomaly models after TensorRT optimization. Step up to Orin NX when you need Transformer-class models or >8 streams per box.

What a real-time pipeline looks like under the hood

A production pipeline is seven stages, not one model call. Missing any stage shows up as false alarms, latency spikes, or storage bills.

1. Ingest. RTSP pull (ONVIF Profile S cameras) or WebRTC for newer stacks. Use GStreamer or NVIDIA DeepStream nvstreammux for multi-stream batching.

2. Decode & pre-process. H.264/H.265 hardware decode on NVDEC or the camera’s SoC. Histogram equalization and resolution crop for the model input.

3. Detection / tracking. YOLOv8 or RT-DETR for objects; ByteTrack for multi-object tracking that survives occlusion; I3D or Video Swin for temporal anomalies.

4. Temporal filter. Require an anomaly signal across 3–5 consecutive frames before firing an alert. This alone removes 40–60% of false positives at a 5–10 ms latency cost.

5. Multi-camera fusion. Correlate events across neighboring cameras via NTP-synced timestamps. Shared anomalies (e.g., a runner crossing two feeds in sequence) get scored higher.

6. Rule & policy engine. Geofences, schedules, access-control events. This is where you convert “person in zone” into “person in restricted zone during off-hours”.

7. Alert dispatch & audit trail. WebSocket to the ops console, push to mobile, write to an immutable audit log. The audit log is what sells you into regulated accounts—skip it and you lose every enterprise deal.

Benchmarks and datasets: what “state of the art” actually means

When a vendor quotes a headline accuracy number, ask which dataset. The five that matter, ranked by difficulty:

UCF-Crime. 1,900 untrimmed videos, 128 hours, 13 real-world anomalies (abuse, robbery, shooting, arson). The hardest public benchmark. Current SOTA: ~84.6% frame-level AUC.

XD-Violence. Large-scale violent behavior, weakly labeled. Current ensemble results: ~88% AUC.

ShanghaiTech. 13 scenes, controlled splits. Easier than UCF-Crime. Top methods: 92–98% AUC.

Avenue. Outdoor traffic anomalies. 85–96% typical.

CUHK Abnormality in Crowds. Pushing, bumping, fighting. 75–95% typical.

Real-world warning: deployment on your own site typically drops AUC by 15–25% because of domain shift. A model that hits 0.90 on ShanghaiTech will land at 0.70–0.75 the first week of site deployment. Budget a retraining sprint on your own labeled data.

The false-positive problem (the only thing operators actually care about)

An untuned model fires a false alarm on 30–70% of events. Past about 30 false alerts per camera per day, human operators start ignoring 40–70% of everything the system raises—including the real events. Every mature deployment spends more engineering on false-positive suppression than on the detector itself.

Five layers actually work, in combination:

1. Adaptive thresholding. Tune per-camera confidence cutoffs on-site. Published adaptive methods report 67% false-positive reduction while keeping >94% true-positive rate.

2. Temporal consensus. Require the anomaly to persist across 3–5 consecutive frames. 40–60% false-positive reduction at negligible latency cost.

3. Multi-camera correlation. Alerts corroborated by a neighboring camera score higher. 70–80% false-positive reduction on regional events.

4. Human-in-the-loop feedback. Let operators flag false positives from the UI; retrain weekly on that feedback. Closes the domain-shift gap in 4–8 weeks.

5. Multi-modal fusion. Combine video with audio, door sensors, access-control events, and (for industrial sites) machine telemetry. Avigilon’s self-learning UMD reports roughly 90% false-alarm reduction in the field.

Buy vs build: third-party platforms you’ll compare against

Any honest custom-build pitch starts by comparing the platform market. If one of these gets you 80% of what you want at 40% of the cost, buy it.

Cost model: a realistic 100-camera deployment

Planning is easier with concrete numbers. Here’s how a 100-camera anomaly-detection deployment lands across the three sourcing paths, at typical 2026 market rates. These ranges assume off-the-shelf anomaly classes (intrusion, loitering, fights, falls). Domain-specific anomalies push custom-build cost up.

Our agent-engineering workflow (LLM-assisted coding, reusable internal libraries for ingest/decoding/tracking/UI) shortens the custom-build timeline versus the baseline industry figures above. The “$50–150k initial” range reflects that, not a low-ball—talk to us with a concrete camera count before you anchor to a number.

Mini case: V.A.L.T in production

V.A.L.T is our long-running video observation and surveillance platform. It sits inside law-enforcement, clinical training and courtroom deployments where the core job is recording, review and rule-based anomaly flagging. The Kazakhstan courtroom deployment is one public example: hundreds of rooms, mandatory audit trails, zero tolerance for missed events.

The architecture that won the contract is the hybrid pattern in this playbook. Edge appliances at each room handle ingest, H.265 encoding, tagging and first-pass anomaly detection. Only events and indexed metadata travel to the central cloud, where supervisors review flagged segments, search across rooms, and export court-admissible packages with tamper-evident logging.

Two lessons carried over to every subsequent surveillance build: invest in the audit trail early because regulated buyers won’t sign without it, and keep the “rules” engine independent of the ML model so non-engineers can add and tune alert policies. Want a similar assessment for your deployment?

A decision framework in five questions

Q1. How many cameras, at how many sites? Below 50 cameras at one site, a commercial VMS like Verkada or Avigilon wins on TCO. Above 200 cameras or across multiple sites with data-residency rules, custom starts to make sense.

Q2. Are your anomalies in the off-the-shelf catalog? Intrusion, loitering, fights, falls, abandoned objects, crowd density—all covered by BriefCam, Avigilon and Verkada. Domain-specific classes (clean-room breach, procedural non-compliance, cockpit behavior, courtroom signals) are not. That’s where custom earns its fee.

Q3. What is your latency budget? Live intervention (robbery, violence) needs <100 ms alerts; forensic review is fine at minutes. Tight latency forces edge or hybrid; forensic can live in the cloud.

Q4. Where does the data legally live? EU AI Act and GDPR effectively kill cloud-first face-recognition-style deployments. If your buyer is a public body or an EU enterprise, default to edge + on-prem storage, detect objects/behavior only.

Q5. Who will retrain the model? Domain shift costs 15–25% AUC on day one. If you don’t have a plan for weekly retraining on site data, pick a vendor that owns that burden or sign a maintenance contract with your custom-build partner.

Five pitfalls that sink deployments

1. Benchmarking on the wrong dataset. Shipping a model that hits 0.95 AUC on ShanghaiTech into a warehouse yard will embarrass you. Always fine-tune on 500+ site-specific clips and report real-world AUC, not research numbers.

2. Treating false positives as a post-launch problem. By the time operators tell you alerts are noisy, they’ve already muted notifications. Temporal filtering, adaptive thresholds and multi-camera fusion go into the MVP, not the backlog.

3. Streaming 4K to the cloud for inference. 500 cameras at 4 Mbps is 21 TB/day. Cloud egress at $0.09/GB adds up to six figures a year. Edge inference is not an optimization, it’s a line-item survival move.

4. Ignoring domain shift and seasonal drift. “Loitering” in an outdoor mall has different baselines in July and December. Without scheduled retraining or online learning, alert precision degrades inside three months.

5. Skipping the audit trail. Regulated buyers (healthcare, law enforcement, courts, finance) walk away the moment they see you can’t produce a tamper-evident log of every alert, every override and every model change. Bolt this on in week one of the build, not month six.

KPIs: what to actually measure

Quality KPIs. Precision and recall on site-specific clips (not public benchmarks). Target: precision ≥ 0.9 at recall ≥ 0.85 after three retraining cycles. Track AUC drift weekly.

Business KPIs. Alerts per camera per day (target <10), operator acknowledgement rate (>80%), time-to-first-response on true positives (<60 s), and percentage of real incidents caught by the system vs. discovered later (>85%).

Reliability KPIs. Edge appliance uptime (99.5%+), detection-to-alert latency p95 (<500 ms), and training pipeline health (successful retraining runs/month, >90%).

Privacy, GDPR and the EU AI Act: what you can and can’t ship

Regulation separates “anomaly detection” into two bands, and treats them very differently.

Low-risk (shippable under GDPR and the EU AI Act). Object detection, motion anomaly, behavior anomaly that does not re-identify individuals. Crowd density, loitering by silhouette, intrusion detection, fall and fight detection. Requires standard GDPR notice and a privacy impact assessment.

High-risk or prohibited. Real-time remote biometric identification in public spaces is prohibited for law enforcement under the EU AI Act. Untargeted scraping of CCTV for facial-recognition databases is explicitly banned. Gait recognition and ethnicity inference are “high-risk” AI systems under the Act and demand full conformity assessment.

Practical guidance: design for the low-risk band by default. If a buyer insists on facial recognition, route the request to a compliance-first vendor and scope a separate high-risk engagement. Enforcement is aggressive—the French DPA fined Clearview AI €20M in 2022 for unlawful face-collection practices, and EU regulators have been consistent since.

When machine-learning anomaly detection is the wrong answer

ML is not a universal surveillance upgrade. Three situations where a rule-based system or a plain human operator is a better fit:

Tiny camera counts with dense human coverage. A single parking-lot camera watched by a 24/7 guard doesn’t benefit enough from an ML layer to justify the hardware, licensing and retraining burden.

Zero training data and zero operator bandwidth. ML needs weeks of labeled “normal” footage plus a feedback loop from operators. If both are missing, a well-tuned motion-detection rule beats an under-trained model.

Life-safety with deterministic rules. Fire, smoke and gas detection are better handled by dedicated sensors than by ML on video. Use video as confirmation, not as primary signal.

A minimal DeepStream pipeline (for engineers scoping this)

For the engineers reviewing this playbook, here’s the skeleton of a multi-stream detection pipeline on Jetson. Useful to set hardware expectations during procurement; not production-ready.

# DeepStream reference pipeline, 4 RTSP streams -> YOLOv8 -> tracker -> sink
gst-launch-1.0 \
  nvstreammux name=mux batch-size=4 width=1280 height=720 live-source=1 ! \
  nvinfer config-file-path=/opt/yolov8.txt ! \
  nvtracker ll-lib-file=/opt/libnvds_nvmultiobjecttracker.so ! \
  nvinfer config-file-path=/opt/anomaly_i3d.txt ! \
  nvmultistreamtiler rows=2 columns=2 ! \
  nvvideoconvert ! nvdsosd ! \
  nveglglessink \
  rtspsrc location=rtsp://cam1:554/h264 ! rtph264depay ! h264parse ! nvv4l2decoder ! mux.sink_0 \
  rtspsrc location=rtsp://cam2:554/h264 ! rtph264depay ! h264parse ! nvv4l2decoder ! mux.sink_1 \
  rtspsrc location=rtsp://cam3:554/h264 ! rtph264depay ! h264parse ! nvv4l2decoder ! mux.sink_2 \
  rtspsrc location=rtsp://cam4:554/h264 ! rtph264depay ! h264parse ! nvv4l2decoder ! mux.sink_3

An Orin Nano clears four 1080p streams with a YOLOv8-s detector and an I3D anomaly head after TensorRT optimization. Swap in a Transformer head and you’ll want Orin NX or better.

FAQ

What to read next

Ready to scope ML anomaly detection for your fleet?

The shortlist is clear. Decide the anomaly classes, camera count and data-residency zone first. Pick hybrid as the default architecture, Jetson Orin Nano or NX as the default edge chip, and I3D or Video Swin as the default model family. Invest in the false-positive stack (adaptive thresholds, temporal consensus, multi-camera fusion, operator feedback) from day one, not after launch.

Then run the build-vs-buy math honestly: if you’re under 150 cameras and your anomalies are generic, a commercial VMS probably wins. Above that, or if your anomaly class is domain-specific, custom is cheaper at five-year TCO and gives you model ownership. Fora Soft has been on the custom side of that line for 21 years—bring us your spec and we’ll tell you which side you’re on before you commit a dollar.