Every year, 3 million elderly adults are treated for fall injuries, making it a leading cause of emergency hospitalization. Sentinel AI was built to solve the high false-alarm rates of traditional motion sensors and the invasiveness of wearable devices.
Sentinel AI is a multimodal, privacy-first fall detection system that understands context. Unlike basic cameras, our Hybrid Engine uses a “Tribunal” voting system between three experts: MediaPipe Pose Physics, an XGBoost ML model, and Audio Spectrogram Transformers (AST) to “hear” the impact when visuals are obscured.
To achieve high-fidelity detection without sacrificing speed, we implemented: • Biomechanical Geometry Engine: Uses joint angles to differentiate between “Bending” (hip flexion/straight knees) and “Falling” (rapid vertical velocity spikes). • Asynchronous Backend Hardening: Solved MediaPipe TFLite thread-affinity issues on Windows using a custom factory pattern for stable 30 FPS processing. • Edge-First Privacy: Runs entirely on local hardware—no video ever leaves the premises. • State Hysteresis: A 30-frame sliding window and One-Euro filters eliminate landmark jitter and “random guessing.”
• Frontend: Next.js 14, Shadcn UI, React Query (Real-time Telemetry) • Backend: Python Flask, MediaPipe 0.10.9 (Stabilized), XGBoost • Inference: Local CPU with Zero-Copy Buffers
📂 GitHub Repo: https://github.com/azaynul10/SentinelAI 💬 Join the Community: https://www.facebook.com/groups/bdaiml/
#AI #ComputerVision #HealthTech #ElderlyCare #MachineLearning #MediaPipe #XGBoost #VibeCoding #SentinelAI
