Clusters after training edge model on someone covering the camera, waiting in front, and people passing by and clear hallways (unknown)
Logo
From the mobile app, system status shows clear hallway
From the mobile app, system status shows someone waiting in front
From the mobile app, system status shows someone covering the camera with their hand
From the mobile app, system status monitoring
Integrated camera, motion sensor, and Arduino Uno Q
Integrated camera, motion sensor, and Arduino Uno Q

NightKnight

Inspiration

NightKnight is a portable, edge-AI powered safety device designed specifically for women and solo travelers staying in hotels, dorms, or short-term rentals.

We wanted to build a portable security system for women and solo travellers

By combining embedded systems and edge AI, we created a portable monitoring system that protects users in real time.

What It Does

NightKnight is an AI-powered smart door monitoring system that performs real-time image classification.

The system detects:

A person waiting in front of the door
A blackout or camera obstruction
Normal or unknown movement

If suspicious behavior is detected, the system triggers an alert response.

Key Features

Edge AI inference (no cloud required)
96 × 96 image input model
INT8 quantized deployment
Real-time classification
Portable plug-and-play hardware

How We Built It

Hardware

Arduino Uno Q
Logitech Brio 100 HD webcam
USB-C interface

Software & AI Pipeline

Collect labeled image data
Train classification model in Edge Impulse
Quantize to INT8
Deploy model
Integrate with hardware

Example class labels:

waiting_in_front
blackout
unknown

Model Summary

Component	Description
Input Size	96 × 96
Quantization	INT8
Classes	3
Inference	Real-time

Challenges We Ran Into

Class imbalance causing incorrect predictions
Motion vs image classification confusion
Deployment troubleshooting
Real-time performance optimization

Accomplishments We’re Proud Of

[x] Trained and deployed a quantized edge AI model
[x] Achieved real-time inference
[x] Integrated camera + microcontroller + AI pipeline
[ ] Add infrared support
[ ] Develop mobile app integration

What We Learned

Through NightKnight, we learned:

Balanced datasets dramatically improve performance
Edge optimization reduces computation time
Hardware constraints influence model design

The prediction function can be expressed as:

( y = f(x; \theta) )

Where:

(x) = input image
(\theta) = learned weights
(y) = predicted class

Display version:

$$ y = \arg\max_i P(class_i \mid x) $$

What's Next for NightKnight

Future Improvements

Add infrared camera support
Integrate rechargeable lithium battery
Add video snapshot when triggered
Allow the application to actually call the emergency contact when the user selects that option
Train thermo-based classification model

NightKnight demonstrates how embedded hardware and AI can be combined to create practical, privacy-focused safety solutions.