AI-Enabled Safe Locker with BrainChip Project

 Build a smart locker system that opens only when both the user’s face and voice command match authorized patterns — all processed using low-power neuromorphic AI. This article provides brief information on an AI-enabled safe locker with BrainChip, features, etc. Please refer to our in-depth article on AI Chips Basics.

AI-Enabled Safe Locker with BrainChip

Features

• Face recognition (Vision AI using SNN)
• Wake word detection (Audio SNN)
• Servo-controlled locking mechanism
• Local, low-latency inference (no cloud)
• On-chip learning (for adding new users)

Please refer to a detailed explanation of how AI GPUs Actually Work.

Components Required

Interested in how NVIDIA compares with AMD in AI computing? Read our detailed comparison NVIDIA vs AMD.

Software Setup

1. Install Dependencies

• sudo apt update.
• sudo apt install python3-pip libatlas-base-dev.
• pip3 install akida speechrecognition opencv-python numpy pyserial.

Install Akida SDK:

Download the BrainChip Akida SDK from the official site. Follow their instructions to install the Python SDK and runtime.

AI-Enabled Safe Locker Project Architecture

Step-by-Step Implementation

Step 1: Data Collection & Preprocessing

a. Face Dataset (Images)

Collect 20–30 frontal face images per authorized person using OpenCV:

import cv2
cap = cv2.VideoCapture(0)
for i in range(30):
ret, frame = cap.read()
cv2.imwrite(f”user_face_{i}.jpg”, frame)
cap.release()

b.Voice Samples (Wake Word)

Record your custom phrase (e.g., “Unlock Akida”) using PyAudio or Audacity.

Step 2 : Train SNN Models with Akida

a. Convert Face Classifier to SNN

Use MobileNet or a custom CNN for feature extraction and convert to SNN using akida.Model.

from akida import Model
model = Model(“cnn_model.h5”)
model.quantize()
model_to_akida = model.convert()
model_to_akida.save(“face_model.akd”)

b. Convert Wake Word Classifier

• Use MFCC preprocessing → CNN → SNN
• Convert the audio classifier to an Akida model using the Akida tools.

Step 3: Load Models and Infer

from akida import AkidaModel
face_model = AkidaModel(“face_model.akd”)
audio_model = AkidaModel(“wake_model.akd”)

Audio Inference (Wake Word)

def is_wake_word(audio):
prediction = audio_model.predict(audio)
return prediction == “unlock_akida”

Face Inference (Real-Time Face Match)

def is_authorized_face(frame):
face = detect_and_crop_face(frame)
prediction = face_model.predict(face)
return prediction == “authorized_user”

Control the Servo Lock

import RPi.GPIO as GPIO
import time
servo_pin = 17
GPIO.setmode(GPIO.BCM)
GPIO.setup(servo_pin, GPIO.OUT)
servo = GPIO.PWM(servo_pin, 50)
servo.start(0)
def open_locker():
servo.ChangeDutyCycle(7.5) # Adjust as per lock
time.sleep(1)
servo.ChangeDutyCycle(0)
def close_locker():
servo.ChangeDutyCycle(2.5)
time.sleep(1)
servo.ChangeDutyCycle(0)

Step 5: Integration Logic

import cv2
import speech_recognition as sr
cam = cv2.VideoCapture(0)
while True:
# Wake word check
audio = record_audio_sample()
if not is_wake_word(audio):
continue
# Face check
ret, frame = cam.read()
if is_authorized_face(frame):
open_locker()
print(“Locker opened!”)
else:
print(“Face not recognized.”)

Testing and Validation

• Add a new user using Brainchip Akida’s on-chip learning API.
• Try unlocking with the wrong voice or face → the system should deny access.
• Log each attempt (success/failure) for analytics.

 Try Implementing the above project  and let us know your results..