When Umbrellas Trigger Lockdowns: Solving False Positives in Weapon Detection AI

The Umbrella That Locked Down a Mall

In August 2025, shoppers at Destiny USA mall in Syracuse, New York, experienced a terrifying afternoon. Multiple 911 calls reported a man carrying what appeared to be an AR-15-style rifle. Police swarmed the building. The mall went into lockdown. Families hid in stores. Parents clutched their children.

The weapon? A closed umbrella with a samurai-style handle, slung over a man's shoulder.

This wasn't a failure of human judgment alone. It's a symptom of a much larger problem that plagues AI-powered weapon detection systems worldwide: the inability to distinguish between actual threats and harmless everyday objects that share similar visual characteristics.

The False Positive Epidemic in Security AI

False positives aren't just embarrassing—they're dangerous. Every false alarm erodes trust in security systems. Security teams become desensitized. Real threats get dismissed as "another umbrella incident." And in worst-case scenarios, innocent people get hurt when responders react to perceived threats that don't exist.

The problem stems from how most weapon detection models are trained. Standard datasets focus almost exclusively on positive examples—images containing actual weapons. The model learns what a rifle looks like, but it never learns what a rifle doesn't look like when confronted with visually similar objects.

Common objects that trigger false positives include:

• Closed umbrellas (especially those with straight handles)
• Camera tripods and monopods
• Long tools like pipe wrenches or crowbars
• Sports equipment (hockey sticks, lacrosse sticks)
• Musical instrument cases
• Walking canes and hiking poles

Hard Negative Mining: Teaching AI What Isn't a Weapon

The solution lies in a technique called hard negative mining. Instead of only showing the model examples of weapons, we deliberately include "confuser" objects—items that look similar to weapons but aren't. By forcing the model to distinguish between rifles and umbrellas, between handguns and power drills, we teach it the subtle differences in grip, posture, and context that separate genuine threats from everyday objects.

Introducing Our Rifles vs. Umbrellas Dataset

Today, we're releasing an open-source synthetic dataset specifically designed to address this challenge. The dataset contains a balanced split of approximately 50% rifles (threat class) and 50% umbrellas (confuser class), all captured from realistic CCTV perspectives.

Sample YAML Configuration

path: /path/to/dataset
train: images
val: images

nc: 3
names:
  0: person
  1: rifle
  2: umbrella

Why Synthetic Data?

Creating a real-world dataset of people carrying rifles in public spaces is, for obvious reasons, impossible. You can't stage photoshoots of people walking through malls with weapons. You can't capture authentic security footage of potential mass shooters for training data. And even if you could, the privacy and ethical implications would be insurmountable.

Synthetic data solves this. Every image in our dataset is computer-generated. No real individuals are depicted. No privacy laws are violated. No ethical boundaries are crossed. And importantly, we have complete control over the scenarios we create—allowing us to generate the exact edge cases that real-world datasets lack.

Use Cases

Download the Dataset

The Bigger Picture

The Destiny USA incident ended without injury—this time. But as AI-powered security systems become more prevalent in schools, malls, airports, and public spaces, the cost of false positives will only grow. Each false alarm wastes resources, traumatizes communities, and chips away at public trust in these systems.

Building better weapon detection AI isn't just about finding more examples of weapons. It's about teaching models the nuanced differences between threats and everyday objects. It's about hard negative mining, synthetic data generation, and rigorous evaluation on real-world edge cases.

Because when an umbrella looks like a rifle to your AI, it's not a software bug—it's a training data problem. And that's a problem we can solve.