By /

Back to: Advanced Physical Security Integration (APSI)

0

Lesson 8.1: The AI Stack (Computer Vision Fundamentals)

Module: 8 – AI & Advanced Analytics Prerequisites: Lesson 3.1 (Camera Hardware) & Lesson 7.4 (Cloud/Edge) Estimated Time: 45–60 Minutes

1. Learning Objectives

By the end of this lesson, you will be able to:

  • Deconstruct the “AI Stack” into its three layers: Capture, Inference, and Action.
  • Differentiate between Pixel Motion Detection (Old School) and Object Classification (New School).
  • Define “Training” vs. “Inference” and why the camera only does one of them.
  • Compare Edge AI (On-Camera) vs. Server AI (On-Premise) vs. Cloud AI.

2. The Evolution: From “Motion” to “Meaning”

To understand AI, you must understand what we are replacing.

  • Pixel Motion (The Old Way):
    • How it worked: The camera compared Frame 1 to Frame 2. If 50% of the pixels changed color, it screamed “ALARM!”
    • The Failure: Trees blowing in the wind, shadows moving, or headlights passing by all change pixels. This caused 95% False Alarm rates.
  • Computer Vision / AI (The New Way):
    • How it works: The camera doesn’t look at pixels; it looks for Shapes and Patterns.
    • The Result: It ignores the moving tree because it knows: “That is a tree, not a human.”

3. The AI Stack: How it Works

AI isn’t magic; it’s a pipeline.

Layer 1: The Model (The Brain)

  • Deep Learning / CNN (Convolutional Neural Network): This is the software brain.
  • Training: You feed a supercomputer 1,000,000 photos of a human. It learns what a “human” looks like (two legs, head, torso). This happens at the factory.
  • Inference: You install this “trained brain” onto the camera. The camera looks at live video and guesses: “I am 98% sure that shape is a Human.”

Layer 2: The Hardware (The Muscle)

  • GPU (Graphics Processing Unit): Massive power, usually in a server (NVIDIA). Used for heavy lifting (Face Recognition).
  • NPU (Neural Processing Unit): A tiny, efficient chip inside the camera. Used for basic tasks (Human/Vehicle detection).

Layer 3: The Application (The Value)

  • What do we do with the data?
  • Security: “Alert guard if Human is in Zone A.”
  • Business: “Count how many Humans entered the store (Retail Analytics).”

4. Deployment Architectures: Where does the AI live?

This is the most expensive decision an integrator makes.

A. Edge AI (On Camera)

  • Location: The analytics run inside the camera chipset (SoC).
  • Pros: Fast, Cheap (no server needed), Low Bandwidth (only sends alerts, not video).
  • Cons: Limited brainpower. Can detect “Human” or “Car,” but maybe not “Steve Smith” or “2015 Toyota Camry.”

B. Server AI (On-Premise)

  • Location: All video is streamed to a big server with NVIDIA GPUs in the closet.
  • Pros: Smarter. Can do Facial Recognition, License Plate Recognition (LPR), and Weapon Detection simultaneously.
  • Cons: Expensive. Requires $5,000+ servers and consumes massive electricity.

C. Cloud AI

  • Location: Video is sent to Amazon/Microsoft/Google servers.
  • Pros: Infinite intelligence. Can search for “Man in red shirt” across 1,000 hours of video in seconds.
  • Cons: High recurring monthly fees. Requires massive internet upload speed.

5. Key Terminology for Integrators

  • Classification: Identifying what an object is (Human, Vehicle, Animal).
  • Attribute Extraction: Identifying details about the object (Red Shirt, Hat, Backpack, White Truck).
  • Behavioral Analytics: Identifying actions (Loitering, Falling Down, Fighting/Fast Movement).
  • False Positive: The alarm went off, but nothing was there (e.g., a scarecrow detected as a human).
  • False Negative: The thief walked right past the camera, and it failed to alert. (This is worse).