Echo Alert

We don't watch for danger. We listen for danger.

Comment

Inspiration

Modern surveillance systems have a critical flaw: they are passive. They record thousands of hours of video, but they don't "understand" what is happening until a human is monitoring the footage continuously..... often too late. We realized that in emergencies like active shooter situations, car crashes, or assaults, sound travels faster than visual confirmation. A scream or a gunshot is an immediate indicator of danger that a camera might miss if the event is out of frame. We wanted to build a system that doesn't just watch, but listens and reacts in real-time, turning passive surveillance into active protection.

What it does

EchoAlert is an Edge-AI powered acoustic surveillance system.

  1. Listens: It continuously monitors audio using a Raspberry Pi.
  2. Detects: It identifies specific dangerous sound signatures (Gunshots, Screams, Explosions, Car Crashes) locally on the device using a custom AI pipeline.
  3. Alerts: Within seconds of detection, it pushes an alert to a central Command Dashboard and a Mobile App for security personnel.
  4. Verifies: It automatically captures and uploads a video clip of the event for immediate evidence, while filtering out non-emergencies to protect privacy.

How I built it

We built a full IoT-to-Cloud pipeline:

  • Hardware: Raspberry Pi 4 Model B acting as the edge node.
  • AI/ML: We used Custom Trained Model and TensorFlow Lite to run a quantized YAMNet model (a deep neural network
    for audio event classification).
  • The "Secret Sauce":
    Raw model predictions were noisy, so we implemented a Few-Shot Learning "Memory Bank". We extract 1024-dimensional embedding vectors from the model and compare them using "Cosine Similarity"
    against known danger signatures. This drastically reduced false positives.
  • Backend: A Python Flask server handles data aggregation, video uploads (Cloudinary), and logs events to Firebase.
  • Frontend: Dashboard: Built with React.js + Vite and Google Maps API for a real-time "Command Center" view. Mobile App: Built with React Native + Expo for field responders, featuring role-based alerts (e.g., Firefighters get
    fire alerts, Police get all alerts, etc).

Challenges I ran into

       1.  **Reducing False Positives**: 
             Initially, the model would confuse loud claps with gunshots. We solved this by moving from simple                            
             classification to Embedding comparison, which acts like a "digital fingerprint" match for sound.
       2.  **Latency vs. Accuracy**: 
            Running deep learning on a Raspberry Pi is heavy. We had to optimize the audio  preprocessing 
            (fasr downsampling from 48kHz to 16kHz) and use TFLite to get inference times under 200ms.
       3.  **Network Complexity**:
            synchronizing the Edge device, local backend, and cloud dashboard required robust networking. 
            We implemented a dual-stack approach (local API for speed, Firebase for remote reliability).

Accomplishments that I'm proud of

  • Real-time Inference: Achieving reliable detection on a Raspberry Pi with minimal latency (<200ms).
  • Privacy First: Implementing local processing so no raw audio is ever streamed to the cloud, setting a new
    standard for surveillance privacy.
  • Memory Bank System: Creating a flexible system that can learn new sounds without needing to retrain the entire
    neural network, making it adaptable to different environments.
  • Full-Stack Integration: Successfully connecting hardware, web, and mobile platforms into a cohesive ecosystem.

What I learned

  • Edge AI Optimization: I gained deep insights into optimizing Python code for ARM processors, specifically using
    NumPy vectorization for efficient audio processing.
  • Audio Engineering: Understanding the nuances of Mel Spectrograms and sample rates was critical for fine-tuning
    the model's accuracy.
  • Resilient System Design: Learning to build a fault-tolerant IoT system capable of handling network interruptions
    and recovering gracefully.

What's next for Echo Alert

  • Sound Triangulation: Implementing a multi-device setup to pinpoint the exact 3D location of a sound source.
  • Drone Integration: Automating the deployment of camera drones to the coordinates of a detected event (e.g., a gunshot) for rapid visual assessment.
  • Mesh Networking: Developing a mesh network capability to allow devices to communicate directly, ensuring
    system functionality even during internet outages.
Share this project:

Updates