Background:
Angel Protection delivers advanced security solutions for educational environments, leveraging object‑detection AI to identify threats in real time. Facing diverse camera setups across schools—with varying resolutions and distortion—and lacking a validated dataset, they needed a scalable, cost‑effective pipeline to detect firearms with minimal false positives and deliver actionable alerts.
Before:
- No existing dataset for gun detection, requiring manual image validation and labeling.
- Relied on Ultralytics (YOLO) with restrictive AGPL-3.0 licensing, complicating commercial deployment.
- Basic weapon-detection capabilities with limited accuracy.
- Hardware costs were prohibitive due to inefficient processing architectures.
- Camera discovery and provisioning relied on error-prone bash scripts.
- Built a custom, license-free detection system that processes feeds from 100+ cameras with ~1-second latency.
- Reduced hardware costs by 90% through a distributed network of small machines.
- Achieved 97.3% accuracy in firearm detection with near-zero false positives using a proprietary YOLO-based model.
- Created a massive image dataset of weapons/scenarios, curated via AI-assisted validation.
- Implemented motion-activated scanning and unwanted-person alerts, cutting response time by 65%.
- Automated camera discovery and based provisioning across many school districts.
Process:
- Developed a distributed architecture: Edge devices handle motion detection, while centralized GPU nodes process inference.
- Trained models on domain-specific data using synthetic augmentations for lighting/angle variations.
- Optimized inference for GPUs via TensorRT, processing 100s of 4k cameras per GPU.
- Creating a commercial friendly python package for Computer Vision (angelcv), increased authroity.
- Implemented camera-agnostic discovery functionality.
Gradient Insight’s solution transformed our approach. We now process 10x more feeds at 1/10th the cost while detecting threats traditional systems miss.
