Detection Best Practices

This guide covers proven patterns for designing effective Anava detections, common mistakes to avoid, and strategies for different use cases.

Core Principles

1. Single Responsibility

Each skill should do one thing well.

Good	Avoid
"Weapon Detection"	"Security Everything"
"PPE Compliance"	"Safety and Security Combined"
"Queue Monitor"	"All Operations Tracking"

Why: Focused skills have:

• Better accuracy
• Clearer prompts
• Easier troubleshooting
• More actionable alerts

2. Appropriate Triggers

Match triggers to your detection needs.

3. Contextual Prompts

Include relevant context in your prompts.

Basic (Less Effective):

Detect people in this image.

Contextual (More Effective):

You are monitoring a warehouse loading dock. During business hours
(7am-5pm), employees and delivery drivers with visible badges are
authorized. After hours, the area should be empty.

Analyze this image and determine if anyone present appears
unauthorized based on location, attire, and behavior.

Detection Patterns by Use Case

Security: Intrusion Detection

Skill: After-Hours Intrusion
Prompts:
  Context: "Secured facility, empty after hours"
  Focus: "Unauthorized presence, entry attempts"
Objects:
  - Person (stateful)
  - Open Door (stateful)
Questions:
  - "Is anyone present?" (bool)
  - "Signs of forced entry?" (bool)

Profile:
  Trigger: Motion or AOAS Person
  Schedule: Outside business hours
  TTS: Enabled with warning message

Security: Weapon Detection

Skill: Weapon Detection
Prompts:
  Context: "Public area requiring safety monitoring"
  Focus: "Visible weapons of any type"
Objects:
  - Weapon (stateful, high priority)
  - Firearm (stateful)
  - Knife (stateful)
Questions:
  - "Type of weapon detected?" (string)
  - "Threat level?" (string: none/low/medium/high/critical)

Profile:
  Trigger: AOAS Person or Motion
  Schedule: 24/7
  Confirmation Threshold: 80%
  TTS: Immediate warning

Safety: PPE Compliance

Skill: PPE Compliance
Prompts:
  Context: "Construction site requiring hard hat and vest"
  Focus: "Missing safety equipment"
Objects:
  - Hard Hat (stateful)
  - Safety Vest (stateful)
  - Safety Glasses (stateful)
Questions:
  - "All PPE present?" (bool)
  - "Missing items?" (set: hard hat, vest, glasses, none)

Profile:
  Trigger: AOAS Person
  Schedule: Work hours only
  TTS: Friendly reminder if PPE missing

Operations: Queue Monitoring

Skill: Queue Length Monitor
Prompts:
  Context: "Retail checkout area"
  Focus: "Queue length and wait times"
Objects:
  - Person (count only, not stateful)
Questions:
  - "Number in queue?" (int)
  - "Estimated wait?" (string)
  - "Queue acceptable?" (bool, threshold: ≤8 people)

Profile:
  Trigger: Schedule (every 60 seconds)
  Schedule: Store hours
  TTS: None (operations only)

Operations: Delivery Monitoring

Skill: Delivery Verification
Prompts:
  Context: "Receiving dock for verified deliveries"
  Focus: "Package handling and verification"
Objects:
  - Person (stateful)
  - Package (stateful)
  - Vehicle (stateful)
Questions:
  - "Is delivery in progress?" (bool)
  - "Proper handling observed?" (bool)

Profile:
  Trigger: AOAS Vehicle + Manual
  Schedule: Receiving hours

Anti-Patterns to Avoid

Anti-Pattern 1: Kitchen Sink Skill

Problem: Skill tries to detect everything.

# DON'T DO THIS
Skill: All Security
Objects:
  - Person
  - Vehicle
  - Weapon
  - Fire
  - Package
  - Animal
  - Badge
  - Uniform
  # 20+ more...

Why It Fails:

• Prompts become unfocused
• Many false positives
• Hard to tune
• Slow analysis

Solution: Create separate focused skills.

Anti-Pattern 2: Vague Prompts

Problem: Prompts lack specific guidance.

# DON'T DO THIS
Prompt: "Look at this image and tell me what's happening"

Why It Fails:

• Inconsistent responses
• Misses important details
• Hard to integrate with VMS

Solution: Specific, structured prompts with clear outputs.

Anti-Pattern 3: Wrong Trigger Type

Problem: Using Motion when Object Analytics available.

Scenario	Bad Choice	Good Choice
Person detection	Motion	AOAS Person
Vehicle detection	Motion	AOAS Vehicle
Perimeter breach	Motion	AOAS Line Crossing

Why It Fails:

• More false positives
• Higher costs
• More noise to filter

Solution: Use most specific trigger available.

Anti-Pattern 4: No Pre-filter

Problem: Every trigger goes to full analysis.

Why It Fails:

• High costs in busy areas
• Unnecessary processing
• Longer response times

Solution: Enable pre-filter for high-traffic scenarios.

Anti-Pattern 5: All Objects Stateful

Problem: Every object emits ONVIF events.

# DON'T DO THIS
Objects:
  - Person (stateful)
  - Bag (stateful)
  - Chair (stateful)
  - Table (stateful)
  # VMS flooded with events!

Why It Fails:

• VMS overwhelmed
• Important events lost in noise
• Action rules become complex

Solution: Only make actionable objects stateful.

Skill Organization Strategies

By Location

Skills:
├── Entrance - General Security
├── Parking - Vehicle Monitoring
├── Warehouse - Safety Compliance
└── Office - Access Control

By Priority

Skills:
├── Critical: Weapon Detection
├── High: Intrusion Detection
├── Medium: PPE Compliance
└── Low: Queue Monitoring

By Function

Skills:
├── Security
│   ├── Weapon Detection
│   └── Intrusion Detection
├── Safety
│   ├── PPE Compliance
│   └── Hazard Detection
└── Operations
    ├── Queue Monitoring
    └── Occupancy Tracking

Tuning Decision Tree

When detections aren't working well:

Performance Optimization

Reduce Analysis Costs

Strategy	Impact	When to Use
AOAS triggers	High	Camera supports AOAS
Pre-filter	High	High-traffic areas
Schedule limits	Medium	Time-bound needs
Lower resolution	Low	Basic detection only

Reduce False Positives

Strategy	Impact	Trade-off
Specific triggers	High	May miss edge cases
Higher confidence	Medium	May miss true positives
Better prompts	High	Requires iteration
Pre-filter	High	Adds latency

Reduce Response Time

Strategy	Impact
Skip pre-filter (for critical)	Faster but more cost
Lower resolution	Faster upload/process
Simpler prompts	Faster analysis

Testing Your Detections

Test Checklist

• True positives detected correctly
• True negatives not triggering
• Edge cases handled
• ONVIF events emitting
• VMS rules triggering
• TTS playing (if configured)
• Performance acceptable

Test Methodology

1. Baseline Test

   • Normal activity, no detection
   • Verify no false positives

2. Positive Test

   • Simulate target scenario
   • Verify correct detection

3. Edge Case Test

   • Partial visibility
   • Poor lighting
   • Multiple objects

4. Stress Test

   • High activity period
   • Multiple simultaneous events

Related Topics

• Detections Overview - Understanding the Detection model
• Tuning & Optimization - Reduce false positives
• Skills Reference - Complete skills documentation
• Profiles Reference - Trigger configuration