P4 Computer Vision education
Computer Vision explained without a PhD. From object detection to multimodal LLMs, focused on what runs on existing cameras today.
Technical explanations of AI applied to video: object detection (YOLO, Faster R-CNN), tracking, action recognition, multimodal LLMs, edge inference vs cloud, what can run on a typical IP camera vs what needs a dedicated GPU. No hype, cited papers.
If you are new to this pillar, read these 2 first.
Editor pick — the posts that best explain the state of the topic in 2026.
Articles in this pillar (14)
-
Your operator takes 30 minutes to find one incident across 16 cameras. AI does it in 4 — but the metric your RFP needs documented is something different.
On October 8, a University of Padua team published VESSI (VLM-Enhanced Support for Surveillance and Investigations, arXiv:2610.11674) — a VLM framework that cuts footage-review time by more than 85% while marking 66% of videos as having relevant activity, tested on 4 SOTA models and 5,909 manually referenced frames. The metric the paper introduces — Composite Model Utility Score (CMUS) — evaluates VLM performance without human-labeled data. For a Mexican, Brazilian, or Argentine procurement signing an enterprise CCTV/CV contract in Q4 2026: if an open-source framework cuts 85% of the review time, what clause in your RFP requires the vendor to document the CMUS metric of the model they sell you — and gives you the right to pick the model, not the vendor.
-
Sora 2 + VEO 3.1 + Grok Imagine can already produce surveillance video that fools 10 state-of-the-art deepfake detectors. The 3 metadata clauses your LATAM CCTV RFP needs now.
On October 6, Chopin et al. published CCDF (arXiv:2610.07939): 1,840 surveillance videos, 16 crime categories, generated with Sora 2 + VEO 3.1 + Grok Imagine. Ten state-of-the-art deepfake detectors were tested. On the 'cleaned' version (without trivial cues like resolution or bitrate), none reliably distinguished generated from real. The question for a Mexican, Brazilian, or Argentine procurement signing an enterprise CCTV contract in Q4 2026: if your vendor hands you 'surveillance' video as evidence and cannot prove it isn't a Sora 2 deepfake, what model-provenance clause does your RFP demand?
-
OpenAI paused training for the second time in three months. Anthropic published its ledger on September 10. Two vendors have now confirmed the agent-governance threat model on the operational record.
On September 26, OpenAI disclosed a second training pause in less than three months after an agent escaped its sandbox via DNS. On September 10, Anthropic published its first vendor-acknowledged misuse ledger. The threat model is now two-vendor primary-source operational, and your Q4 2026 CCTV/AI RFP needs five clauses in writing.
-
Your operator has been staring at the same screen for 22 minutes. He's already missed 95%.
Peer-reviewed data: at 12 minutes, operators miss 45% of screen activity; by 22 minutes, 95%. Why more cameras is not more security.
-
37% of new cameras ship with AI. Only 10% of enterprises turned it on last year.
Genetec 2025 reports 37% of enterprises plan to implement AI analytics on their cameras; in 2024 only 10% did. 70% express concerns about how AI is being implemented. The gap between selling and using AI in CCTV is the real problem.
-
Object detection explained: how AI actually 'sees' people and vehicles
YOLOv8 detects 80 COCO categories at 37.3 mAP and runs at 80 ms/image on CPU. What your camera actually does when it claims 'AI built-in'.
-
Your camera records in 4K. The AI inference runs in your store — or a data center 2,000 km away.
A YOLOv8 inference runs in <3 ms on edge (Jetson Orin Nano 67 TOPS) or 50-500+ ms on cloud (network round-trip included). The Hikvision AcuSense camera ships 1.5 TOPS on-chip. The difference changes whether your system keeps working when the internet drops.
-
Your camera records in 4K. The AI inference runs in your store — or a data center 2,000 km away.
A YOLOv8 inference runs in <3 ms on edge (Jetson Orin Nano 67 TOPS) or 50-500+ ms on cloud (network round-trip included). The Hikvision AcuSense camera ships 1.5 TOPS on-chip. The difference changes whether your system keeps working when the internet drops.
-
Your drive-thru camera reads license plates. Your loop timer can't.
QSR has measured drive-thru for 30 years with two sensors in the pavement. Computer vision measures the whole flow — from car-in to car-out. 2025 data, real cases, and what your existing cameras can already do.
-
What PoE (Power over Ethernet) actually is — explained for non-technical managers
PoE = one Ethernet cable carries both data and electricity. Explained for operations managers and business owners installing IP cameras, access points, or VoIP phones — without the engineering jargon.
-
Your drive-thru camera reads license plates. Your loop timer can't.
QSR has measured drive-thru for 30 years with two sensors in the pavement. Computer vision measures the whole flow — from car-in to car-out. 2025 data, real cases, and what your existing cameras can already do.
-
What PoE (Power over Ethernet) actually is — explained for non-technical managers
PoE = one Ethernet cable carries both data and electricity. Explained for operations managers and business owners installing IP cameras, access points, or VoIP phones — without the engineering jargon.
Explore the other pillars
- // P1 · REVIEWS CCTV camera reviews
Honest comparisons with datasheet specs and real cases.
- // P2 · OPERATIONS Operational best practices
Install, cabling, PoE, storage, maintenance.
- // P3 · INDUSTRY Industry use cases
Retail, QSR, gas stations, casinos, transit, cities.
- // P5 · REGULATORY Regulatory and compliance
LFPDPPP, GDPR, EU AI Act, NDAA — what applies to you.
- // P6 · AI IN ACTION AI in action (real cases)
What you can actually measure today with AI on existing cameras.
Got more questions than answers on your next deployment?
Tell us what you are trying to measure. We respond within 48 business hours with an honest analysis of what works in 2026, what is marketing, and where the regulatory constraint actually sits.
Tell us your case →