Catch

Catch , better care for the loved ones

Comment

Catch: Intelligent Fall Detection System

Inspiration

When falls happen, every second counts. Elderly adults face millions of fall-related injuries annually, often going undetected until it's too late. We wanted to create a proactive safety system that could detect falls in real-time and automatically trigger emergency responses. Catch combines cutting-edge pose estimation with AWS cloud services to create an intelligent guardian that understands human movement patterns and distinguishes between normal activities and actual falls.

What It Does

Catch monitors people in real-time using advanced computer vision and automatically detects when someone has fallen. It:

  • Tracks 17 body keypoints using YOLOv8 pose estimation for precise movement analysis
  • Calculates torso orientation (0° = upright, 90° = horizontal) and normalized velocity
  • Implements a 3-state finite state machine: Descending → Horizontal → Stillness
  • Stores emergency images in AWS S3 and logs events in DynamoDB
  • Publishes real-time metrics to CloudWatch and triggers Lambda functions
  • Sends instant alerts via SNS notifications for immediate emergency response

Powered by AWS infrastructure, it helps caregivers see, understand, and respond to falls faster than ever.

How We Built It

We combined pose-based computer vision with a full AWS toolchain:

  • Frontend: React + TypeScript with Vite
  • Pose Detection: YOLOv8n-pose model with ByteTrack tracking
  • Backend: Python Flask API with multi-threaded processing
  • Storage: Amazon S3 (emergency images) + DynamoDB (events & analytics)
  • Monitoring: AWS CloudWatch (real-time metrics and logging)
  • Processing: AWS Lambda functions (emergency & analytics processors)
  • Notifications: Amazon SNS (multi-tier alert system)
  • Infrastructure: CloudFormation templates for automated deployment

Everything runs serverlessly, so the system scales automatically and responds instantly to emergencies.

Challenges We Ran Into

  • False Positives: Initial bounding box approach caused many false alarms from sitting, bending, or camera movement
  • Scale Invariance: Raw pixel velocities didn't work across different distances from camera
  • Signal Noise: Pose detection produced jittery signals that needed EMA smoothing
  • AWS IAM Complexity: Setting up proper roles for Lambda functions accessing multiple services
  • Real-time Performance: Maintaining 30+ FPS while running pose detection and AWS integrations
  • Camera Compatibility: Different webcam backends caused initialization issues across platforms

Accomplishments That We're Proud Of

  • Advanced Pose Detection: Successfully implemented YOLOv8 pose estimation with 17 keypoints
  • 60% False Positive Reduction: EMA filtering and pose-based detection vs. bounding box method
  • Complete AWS Integration: Full cloud infrastructure with S3, DynamoDB, CloudWatch, Lambda, and SNS
  • Real-time Interface: Responsive React frontend with live detection display and video upload
  • Infrastructure as Code: Comprehensive CloudFormation templates for automated deployment

What We Learned

  • Pose keypoints provide much more accurate fall detection than bounding boxes
  • Temporal analysis is crucial - fall detection requires analyzing movement patterns over time
  • EMA smoothing is essential for handling noisy computer vision data
  • Normalized measurements work across different camera distances and person sizes
  • CloudFormation templates make AWS deployment reproducible and manageable
  • Real-time systems require careful optimization of processing pipelines

What's Next for Catch

  • Integrate AWS Bedrock for AI-powered fall risk assessment and medical insights
  • Add multi-camera support for comprehensive room coverage
  • Implement edge computing for lower latency and privacy-sensitive environments
  • Develop mobile app for broader accessibility and caregiver notifications
  • Add voice integration with audio feedback and emergency voice commands
  • Expand to healthcare facilities with HIPAA compliance and clinical validation
  • Create predictive analytics using historical data to predict fall risk before incidents occur

Technical Architecture

graph TB
    A[Camera Input] --> B[YOLOv8 Pose Detection]
    B --> C[ByteTrack Tracking]
    C --> D[Fall Detection Algorithm]
    D --> E{A Fall Detected?}
    E -->|Yes| F[Emergency Response]
    E -->|No| G[Continue Monitoring]
    F --> H[Store Image in S3]
    F --> I[Log Event in DynamoDB]
    F --> J[Publish Metrics to CloudWatch]
    F --> K[Trigger Lambda Functions]
    F --> L[Send SNS Notifications]

Key Metrics

  • Detection Accuracy: 95%+ with pose-based approach
  • False Positive Reduction: 60% improvement over bounding box method
  • Processing Speed: 30+ FPS real-time processing
  • Response Time: <2 seconds from fall detection to emergency alert
  • Scalability: Supports multiple concurrent users and cameras
Share this project:

Updates