% Load your video data and labels
% Assume you have a cell array 'videoData' containing video sequences and
% a corresponding categorical array 'labels' containing class labels.
% Set the input and output paths
inputPath = "E:\PHB\presentation back up\datasets\violence detection";
outputPath = "E:\PHB\presentation back up\new action\action5\frames";
% Set the number of frames per clip
numFramesPerClip = 64; % Adjust as needed
% Set the subfolders as the categories
categories = dir(inputPath);
categories = {categories(3:end).name};
numClasses = length(categories);
% Create output folder if it doesn't exist
if ~exist(outputPath, 'dir')
mkdir(outputPath);
end
% Loop over categories
for i = 1:numel(categories)
category = categories{i};
categoryPath = fullfile(inputPath, category);
% Create subfolder in the output path for the current category
outputCategoryPath = fullfile(outputPath, category);
if ~exist(outputCategoryPath, 'dir')
mkdir(outputCategoryPath);
end
videos = dir(fullfile(categoryPath, '*.mp4'));
% Set the label for the current category
ll = i * ones(numel(videos), 1);
% Loop over the videos in the current category
for j = 1:numel(videos)
videoPath = fullfile(categoryPath, videos(j).name);
[~, videoName, ~] = fileparts(videos(j).name); % Extract video name
disp(['Processing video: ', videoPath]);
v = VideoReader(videoPath);
% Extract frames
frameIdx = 1;
while hasFrame(v) && frameIdx <= numFramesPerClip
frame = readFrame(v);
% Resize frame
resizedFrame = imresize(frame, [224, 224]);
% Save the frame with a unique name
frameFileName = sprintf('%s_frame_%04d.jpg', videoName, frameIdx);
imwrite(resizedFrame, fullfile(outputCategoryPath, frameFileName));
frameIdx = frameIdx + 1;
end
end
end
% Define the input size based on your frame size
inputSize = [224 224 3];
% Define the DCNN architecture
layers = [
imageInputLayer(inputSize, 'Name', 'input', 'Normalization', 'none')
convolution2dLayer(3, 32, 'Padding', 'same', 'Name', 'conv1')
reluLayer('Name', 'relu1')
maxPooling2dLayer(2, 'Stride', 2, 'Name', 'pool1')
convolution2dLayer(3, 64, 'Padding', 'same', 'Name', 'conv2')
reluLayer('Name', 'relu2')
maxPooling2dLayer(2, 'Stride', 2, 'Name', 'pool2')
fullyConnectedLayer(256, 'Name', 'fc1')
reluLayer('Name', 'relu3')
dropoutLayer(0.5, 'Name', 'drop1')
fullyConnectedLayer(2, 'Name', 'output')
softmaxLayer('Name', 'softmax')
classificationLayer('Name', 'classoutput')
];
% Define LSTM layer
hiddenUnits = 100; % Adjust based on your requirements
lstmLayerObj = lstmLayer(hiddenUnits, 'OutputMode', 'last', 'Name', 'lstm_layer');
% Define the rest of the layers for combining CNN and LSTM
combinedLayers = [
fullyConnectedLayer(100, 'Name', 'fc_final')
softmaxLayer('Name', 'softmax_final')
classificationLayer('Name', 'classOutput')
];
% Set training options
options = trainingOptions('sgdm', ...
'MaxEpochs', 10, ...
'MiniBatchSize', 32, ...
'Shuffle', 'every-epoch', ...
'Verbose', true, ...
'Plots', 'training-progress');
% Convert video data to sequences of image data
sequences = cell(numClasses, 1);
for i = 1:numClasses
category = categories{i};
categoryPath = fullfile(outputPath, category);
frames = dir(fullfile(categoryPath, '*.jpg'));
numFrames = min(numFramesPerClip, numel(frames));
currentSequence = zeros(inputSize(1), inputSize(2), inputSize(3), numFrames);
for k = 1:numFrames
frame = imread(fullfile(categoryPath, frames(k).name));
currentSequence(:, :, :, k) = im2double(frame);
end
sequences{i} = currentSequence;
end
% Display the layer names
layerNames = {lgraph.Layers.Name}
% Create a layer graph and add layers to it
lgraph = layerGraph();
lgraph = addLayers(lgraph, layers);
lgraph = addLayers(lgraph, lstmLayerObj);
lgraph = addLayers(lgraph, combinedLayers);
% Connect layers
lgraph = connectLayers(lgraph, 'fc_final', 'softmax');
lgraph = connectLayers(lgraph, 'softmax', 'classOutput');
% Train the model
net = trainNetwork(sequences, categorical(ll), lgraph, options);
