Person Detection That Actually Works: Privacy-First Computer Vision
The Challenge
A security camera startup needed accurate person detection without sending video to the cloud. Their early prototype used cloud-based detection, which raised privacy concerns and had high false-positive rates (alerting for trees, shadows, pets).
The Problems:
- •False positive rate: 15% (users received alerts for non-persons)
- •Cloud processing created 3-5 second latency for alerts
- •Privacy concerns: customers did not want video uploaded to cloud
- •High AWS costs: $8/camera/month for cloud inference
The Market Opportunity
Competitors offered cloud-only solutions. A privacy-first camera with fast, accurate person detection could be a major differentiator in the crowded home security market.
The Technical Challenge
Run state-of-the-art person detection on a $30 camera with limited compute power, no GPU, and strict power budget. Achieve accuracy matching cloud-based systems while processing everything locally.
Our Approach
R - Recognize (Weeks 1-3)
We analyzed the existing system and identified root causes of poor performance.
- • Cloud model was too generic (trained on COCO dataset)
- • Not optimized for home security use cases (doorways, driveways)
- • Latency from cloud upload/download killed real-time detection
- • Target hardware: ARM Cortex-A7 processor (no GPU acceleration)
E - Envision (Weeks 4-7)
We designed an edge-first computer vision system.
- • Target: 99%+ accuracy, <1% false positives
- • All processing on-device (privacy-first architecture)
- • Real-time alerts (<500ms from detection to notification)
- • Power budget: <5W total for continuous operation
A - Architect (Weeks 8-12)
We designed a custom person detection pipeline for edge deployment.
- • Model: MobileNetV2 backbone with SSD detection head
- • Training: Fine-tuned on 50K images of home security scenarios
- • Optimization: Quantized to INT8, pruned to 2.8MB model size
- • Multi-stage pipeline: motion detection → person detection → alert
C - Construct (Weeks 13-28)
We built and optimized the edge AI person detection system.
- • Collected training data: 50K labeled images from beta cameras
- • Trained custom model: 99.2% precision, 98.7% recall on validation set
- • Optimized for ARM NEON: 180ms inference on Cortex-A7
- • Integrated with camera firmware and notification system
T - Transmute (Weeks 29-36)
We deployed to production and scaled to thousands of units.
- • Beta testing: 500 cameras in real homes for 6 weeks
- • Collected edge case data: challenging lighting, weather conditions
- • Model retraining: reduced false positives from 2.1% to 0.8%
- • Production deployment: OTA update to all units
O - Operate (Ongoing)
We provide continuous monitoring and model improvements.
- • Monthly analytics: accuracy metrics across 500K+ deployed cameras
- • Quarterly retraining with new edge cases
- • A/B testing of model architecture improvements
- • Performance optimization: reduced inference to 140ms
R - Result (Month 10)
Business impact 4 months after full production deployment.
- • Became #1 best-selling security camera on Amazon
- • 4.7-star rating with 50K+ reviews
- • Scaled from 10K to 500K units shipped
- • Eliminated $4M annually in cloud infrastructure costs
The Solution
Technical Specifications
Model Architecture:
- • MobileNetV2 backbone (pretrained on ImageNet)
- • SSD (Single Shot Detector) head for person detection
- • Input: 320x320 RGB frames at 15 FPS
- • Model size: 2.8MB (quantized to INT8)
Training Data:
- • 50K labeled images from home security cameras
- • Diverse conditions: day/night, rain, snow, fog
- • Hard negatives: trees, shadows, pets, vehicles
- • Augmentation: brightness, contrast, rotation variations
Performance:
- • Accuracy: 99.2% precision, 98.7% recall
- • False positive rate: 0.8% (down from 15%)
- • Inference time: 140ms per frame on ARM Cortex-A7
- • Power consumption: 3.2W total (camera + inference)
Deployment:
- • Runs on ARM Cortex-A7 (no GPU required)
- • TensorFlow Lite with ARM NEON optimizations
- • Multi-stage pipeline: motion → detection → classification
- • Alert latency: <500ms from person detected to notification
Privacy Features:
- • 100% on-device processing (no cloud uploads)
- • Optional local storage only (encrypted SD card)
- • No identifiable video data leaves the device
- • User control: can disable cloud connectivity entirely
The Results
Accuracy
Person detection with near-zero false positives
On-Device
All video processing stays local for privacy
on Amazon
In the home security camera category
Business Impact
- • Became #1 best-selling security camera on Amazon
- • 4.7-star rating with 50,000+ customer reviews
- • Scaled from 10K to 500K units shipped in 12 months
- • Eliminated $4M annually in cloud infrastructure costs
- • Privacy-first positioning became key brand differentiator
- • 94% of reviews mention fast, accurate person detection
"Business Reactor helped us build a computer vision system that outperforms competitors while respecting user privacy. Our customers love that their video never leaves the device. It is our biggest competitive advantage."
— CEO, Security Camera Startup
Key Takeaways
1. Privacy Is a Feature, Not a Trade-Off
By processing everything on-device, we eliminated privacy concerns AND reduced latency AND eliminated cloud costs. Privacy-first architecture can be a win-win-win.
2. Domain-Specific Training Beats Generic Models
Generic models trained on COCO dataset perform poorly on home security use cases. We collected 50K images from real doorbell cameras and achieved 10x better accuracy than off-the-shelf models.
3. False Positives Kill User Trust
Users will tolerate occasional missed detections, but frequent false alarms destroy trust in the system. We prioritized reducing false positives even at the cost of slightly lower recall.
4. Multi-Stage Pipelines Save Power
Running person detection continuously would drain battery. We used a lightweight motion detector to trigger the heavier person detection model only when needed, reducing power consumption by 80%.
5. Edge Cases Define Production Quality
The model worked great in testing but failed in rain, fog, and at night. We deployed to 500 beta homes to discover these edge cases, then retrained. Production AI requires real-world data.
Related Projects
Smart Audio
How we fixed spatial audio calibration and improved customer ratings from 3.9 to 4.8 stars.
Read Case StudyAppliance AI
Predictive maintenance that reduced warranty claims by 30% and saved $630K annually.
Read Case StudyNeed Privacy-First Computer Vision?
Let's discuss how edge AI can help you build products that respect user privacy while delivering superior performance.
Schedule Discovery Call