Tuning & Optimization

This guide covers strategies for reducing false positives, improving detection accuracy, and optimizing your Anava deployment.

Understanding False Positives

False positives occur when Anava detects something that isn't actually there or misclassifies what it sees.

Common Causes

Cause	Example	Solution
Wrong trigger	Motion from shadows	Use AOAS Person trigger
Vague prompts	Unclear what to detect	Add specific context
Low confidence	Uncertain detections pass	Increase threshold
Environmental	Weather, lighting changes	Add pre-filter
Scene complexity	Busy areas with many objects	Narrow focus

Optimization Decision Tree

Strategy 1: Use Better Triggers

The trigger is your first line of defense against false positives.

Trigger Effectiveness

Trigger	False Positive Rate	Best For
Motion	High	General monitoring
AOAS Person	Low	Security, PPE
AOAS Vehicle	Low	Parking, traffic
Digital Input	Very Low	Door/sensor events
Manual	None	Operator-initiated

Migrating from Motion to AOAS

1. Verify camera supports AXIS Object Analytics
2. Configure AOAS scenarios on camera
3. Create new profile with Object trigger
4. Test alongside existing Motion profile
5. Disable Motion profile when satisfied

When to Keep Motion

• Camera doesn't support AOAS
• Detecting objects AOAS doesn't recognize
• Very low-traffic areas where every event matters

Strategy 2: Enable Pre-filtering

Pre-filtering adds a fast check before full analysis.

How Pre-filter Works

Configuring Pre-filter

In your skill's Analysis Configuration:

Pre-filter Criteria:

Human presence in the frame

Pre-filter Prompt:

Is there a person clearly visible in this image? Answer only yes or no.

Pre-filter Best Practices

Scenario	Pre-filter
Security (people)	"Is there a person?"
Vehicle monitoring	"Is there a vehicle?"
PPE compliance	"Is there a person in work area?"
Fire detection	"Is there smoke or flame visible?"

When NOT to Use Pre-filter

• Every trigger is potentially critical (e.g., weapon detection)
• Very low-traffic areas
• Latency is critical

Strategy 3: Adjust Confidence Thresholds

Confidence thresholds control when detections become events.

Understanding Confidence

Level	Meaning	Use
0-50%	Low confidence	Usually filtered out
50-70%	Medium confidence	Review needed
70-90%	High confidence	Likely accurate
90-100%	Very high confidence	Almost certain

Setting Thresholds

In profile settings:

Use Case	Recommended Threshold
Critical (weapons)	80%
Security (intrusion)	85%
Compliance (PPE)	75%
Operations (queues)	70%

Trade-offs

Higher Threshold:
├── ✓ Fewer false positives
├── ✓ Less noise
├── ✗ May miss some detections
└── ✗ Could miss edge cases

Lower Threshold:
├── ✓ Catches more events
├── ✓ Good for critical scenarios
├── ✗ More false positives
└── ✗ More review needed

Strategy 4: Improve Prompts

Better prompts lead to better accuracy.

Prompt Improvement Checklist

• Context included (location, time, expectations)
• Specific objects named
• Clear output format defined
• Common false positives addressed

Before/After Examples

Before (Vague):

Look for people in this image.

After (Specific):

This camera monitors a warehouse loading dock. During business hours
(7am-5pm), employees wear yellow vests and hard hats. After hours,
the area should be empty.

Identify any people and determine if they appear to be authorized
employees based on attire and behavior.

Before (No Context):

Detect weapons.

After (Contextual):

Monitor this school entrance for potential weapons. Consider:
- Firearms of any type
- Knives or bladed weapons
- Blunt weapons (bats, clubs)
- Improvised weapons

Exclude common items like umbrellas, sports equipment (when
contextually appropriate), and work tools.

Strategy 5: Split Skills

Focused skills perform better than broad ones.

Signs You Should Split

• Skill has 10+ objects
• Prompts are long and complex
• Different objects need different contexts
• False positives vary by object type

Splitting Example

Before: One Security skill

• Person, Vehicle, Weapon, Fire, Package, Animal, Badge...

After: Focused skills

• Intrusion Detection (Person, Unauthorized Access)
• Weapon Detection (Weapon, Firearm, Knife)
• Fire Safety (Smoke, Flame, Evacuate)
• Package Monitoring (Package, Delivery)

Measuring Improvement

Key Metrics

Metric	Target	How to Measure
False Positive Rate	Under 10%	Review sessions, count incorrect
Miss Rate	Under 5%	Test with known scenarios
Response Time	Under 5s	Check session timestamps

A/B Testing

1. Create new profile with changes
2. Run alongside existing profile
3. Compare session quality
4. Gradually shift to better config

Session Review Process

1. Navigate to Sessions
2. Filter by profile/skill
3. Review random sample (10-20)
4. Categorize: True Positive, False Positive, Miss
5. Calculate rates

Environment-Specific Tuning

Outdoor Cameras

Challenges:

• Weather (rain, snow, fog)
• Lighting (shadows, sun glare)
• Wildlife (animals triggering)

Solutions:

• Use AOAS with human/vehicle classification
• Add weather context to prompts
• Enable pre-filter for human presence

Indoor Cameras

Challenges:

• Reflections (windows, mirrors)
• Motion (fans, curtains)
• Varying activity levels

Solutions:

• Use AOAS where possible
• Add environmental context
• Schedule-based thresholds

Low-Light Scenarios

Challenges:

• Poor image quality
• IR artifacts
• Thermal noise

Solutions:

• Lower confidence expectations
• Simpler detection goals
• Consider higher resolution

Troubleshooting Specific Issues

"Person" Detections on Objects

Problem: Mannequins, posters, or statues trigger person detection.

Solutions:

1. Add to prompt: "Ignore mannequins, posters, and stationary displays"
2. Use AOAS Person (better at distinguishing)
3. Increase confidence threshold

Motion from Environmental Factors

Problem: Shadows, trees, or lighting changes trigger motion.

Solutions:

1. Switch to AOAS trigger
2. Enable pre-filter with "Is there a person/vehicle?"
3. Adjust camera's motion sensitivity

Inconsistent Results

Problem: Same scenario gives different results.

Solutions:

1. Make prompts more structured
2. Use boolean questions for consistency
3. Review multiple sessions to find patterns

Related Topics

• Detections Overview - Understanding the Detection model
• Confidence Tuning - Detailed threshold configuration
• Learning Mode - Active monitoring setup
• Best Practices - Detection design patterns