AI가 졸린지 아닌지 판별할 수 있는 건에 관하여

This app begins by setting up directories to load images of closed and open eyes, followed by images of yawning and non-yawning faces. These images are then labeled and split into training and testing sets to prepare for building two convolutional neural networks (CNNs). One CNN is trained to recognize whether eyes are open or closed, while the other identifies yawning. With the help of image augmentation and categorical labels, the networks are trained using a stochastic gradient descent with momentum (SGDM) optimizer, ensuring they learn to classify images effectively.

The real magic happens when the trained models are loaded and tested on new images to detect drowsiness. The script employs eye and mouth detectors to locate these features in each image. By processing these regions and feeding them into the respective CNNs, it determines whether the driver's eyes are closed or if they are yawning. It counts these occurrences, calculates a drowsiness score based on a simple formula, and finally, decides if the driver is drowsy. With a blend of image processing, machine learning, and a dash of statistical computation, this project not only showcases the power of AI but also provides a practical solution for enhancing driver safety.

 

%% Set directories for training data
dirClosedEyes = "C:\Users\train\Closed_Eyes";
dirOpenEyes = "C:\Users\train\Open_Eyes";

%% Load images and create labels
[imagesClosed, labelsClosed] = loadImagesAndLabels(dirClosedEyes, 1); % 1 for Closed Eyes
[imagesOpen, labelsOpen] = loadImagesAndLabels(dirOpenEyes, 0); % 0 for Open Eyes

images = cat(4, imagesClosed, imagesOpen);
labels = cat(1, labelsClosed, labelsOpen);

%% Split data into training and testing sets
[idx, numTrain, x_train_eye, y_train_eye, x_test_eye, y_test_eye] = splitData(images, labels, 0.8);

%% Define CNN architecture for eye detection
layers_eye = createCNNLayers();

%% Set training options
options_eye = trainingOptions('sgdm', ...
    'InitialLearnRate', 0.01, ...
    'MaxEpochs', 10, ...
    'Shuffle', 'every-epoch', ...
    'ValidationData', {x_test_eye, y_test_eye}, ...
    'ValidationFrequency', 30, ...
    'Verbose', false, ...
    'Plots', 'training-progress');

%% Train the network
netEye = trainNetwork(x_train_eye, y_train_eye, layers_eye, options_eye);
save('netEyeModel.mat', 'netEye');

%% Evaluate the network
predictedLabels = classify(netEye, x_test_eye);
accuracy = sum(predictedLabels == y_test_eye) / numel(y_test_eye);
fprintf('Accuracy: %.2f%%\n', accuracy * 100);

%% Function to load images and labels
function [images, labels] = loadImagesAndLabels(imageDir, label)
    imageFiles = dir(fullfile(imageDir, '*.png'));
    numImages = length(imageFiles);
    images = zeros(64, 64, 1, numImages, 'uint8');
    labels = repmat(label, numImages, 1);

    for i = 1:numImages
        img = imread(fullfile(imageDir, imageFiles(i).name));
        img = imresize(img, [64, 64]);
        if size(img, 3) == 3
            img = rgb2gray(img);
        end
        images(:, :, 1, i) = img;
    end
end

function [idx, numTrain, x_train, y_train, x_test, y_test] = splitData(images, labels, trainRatio)
    idx = randperm(size(images, 4));
    numTrain = round(trainRatio * numel(labels));
    x_train = images(:, :, :, idx(1:numTrain));
    y_train = labels(idx(1:numTrain));
    x_test = images(:, :, :, idx(numTrain+1:end));
    y_test = labels(idx(numTrain+1:end));

    y_train = categorical(y_train);
    y_test = categorical(y_test);
end

%% Function to create CNN layers
function layers = createCNNLayers()
    layers = [
        imageInputLayer([64 64 1])
        convolution2dLayer(5, 32, 'Padding', 'same')
        batchNormalizationLayer
        reluLayer
        maxPooling2dLayer(2, 'Stride', 2)
        fullyConnectedLayer(2)
        softmaxLayer
        classificationLayer];
end

%% Set directories for yawning data
dirNoYawn = "C:\Users\train2\no_yawn";
dirYawn = "C:\Users\train2\yawn";

%% Load images and create labels
[imagesNoYawn, labelsNoYawn] = loadImagesAndLabels(dirNoYawn, 'no_yawn');
[imagesYawn, labelsYawn] = loadImagesAndLabels(dirYawn, 'yawn');

images = cat(4, imagesNoYawn, imagesYawn);
labels = [labelsNoYawn; labelsYawn];

%% Split data into training and testing sets
[idx, numTrain, x_train_yawn, y_train_yawn, x_test_yawn, y_test_yawn] = splitData(images, labels, 0.8);

%% Define CNN architecture for yawning detection
layers_yawn = createCNNLayers();

%% Set training options
options_yawn = trainingOptions('sgdm', ...
    'InitialLearnRate', 0.01, ...
    'MaxEpochs', 10, ...
    'Shuffle', 'every-epoch', ...
    'ValidationData', {x_test_yawn, y_test_yawn}, ...
    'ValidationFrequency', 30, ...
    'Verbose', true, ...
    'Plots', 'training-progress');

%% Train the network
netYawn = trainNetwork(x_train_yawn, y_train_yawn, layers_yawn, options_yawn);
save('netYawnModel.mat', 'netYawn');

%% Evaluate the network
predictedLabels = classify(netYawn, x_test_yawn);
accuracy = sum(predictedLabels == y_test_yawn) / numel(y_test_yawn);
fprintf('Accuracy: %.2f%%\n', accuracy * 100);

%% Function to detect blinking
function isBlinking = detectBlinking(image, blinkingModel)
    img = preprocessImage(image);
    label = classify(blinkingModel, img);
    isBlinking = (label == "closed");
end

%% Function to detect yawning
function isYawning = detectYawning(image, yawningModel)
    img = preprocessImage(image);
    label = classify(yawningModel, img);
    isYawning = (label == "yawn");
end

%% Function to preprocess images
function img = preprocessImage(image)
    img = imresize(image, [64, 64]);
    if size(img, 3) == 3
        img = rgb2gray(img);
    end
    img = reshape(img, [64, 64, 1, 1]);
end

%% Load trained models
load('netEyeModel.mat', 'netEye');
load('netYawnModel.mat', 'netYawn');

%% Set directory for drowsiness data
imageDir = "C:\Users\Drowsiness_Dataset\Drowsy";
imageFiles = dir(fullfile(imageDir, '*.png'));
images = cell(1, length(imageFiles));

for i = 1:length(imageFiles)
    images{i} = imread(fullfile(imageDir, imageFiles(i).name));
end

%% Eye and mouth detectors
eyeDetector = vision.CascadeObjectDetector('EyePairBig');
mouthDetector = vision.CascadeObjectDetector('Mouth');

closedEyesCount = 0;
yawningCount = 0;

%% Process each image
for i = 0:2000
    image = images{i};
    
    % Eye detection and classification
    bbox = step(eyeDetector, image);
    if ~isempty(bbox)
        croppedEyes = imcrop(image, bbox(1,:));
        if detectBlinking(croppedEyes, netEye)
            closedEyesCount = closedEyesCount + 1;
        end
    end
    
    % Mouth detection and yawning classification
    mouthBbox = step(mouthDetector, image);
    if ~isempty(mouthBbox)
        [~, largestBboxIndex] = max(mouthBbox(:,3).*mouthBbox(:,4));
        croppedMouth = imcrop(image, mouthBbox(largestBboxIndex,:));
        if detectYawning(croppedMouth, netYawn)
            yawningCount = yawningCount + 1;
        end
    end
end

%% Calculate drowsiness score
alpha = 0.6;
beta = 0.12;
drowsinessScore = alpha + beta * (closedEyesCount - 20) + log10(yawningCount + 1);
drowsinessScore = max(0, min(1, drowsinessScore));

%% Determine drowsiness state
threshold = 0.5;
if drowsinessScore > threshold
    disp('Driver is drowsy');
else
    disp('Driver is not drowsy');
end