基于MATLAB的HOG+SVM目标检测案例举例，包含完整的代码及代码讲解

好的，我会为您提供一个基于MATLAB的HOG+SVM目标检测案例，包含完整的代码及代码讲解。

首先，让我们来看一下HOG+SVM目标检测的基本流程：

1. 收集和准备训练数据集，包括目标和非目标数据集。

2. 提取HOG特征，将训练数据集转换为HOG特征向量。

3. 使用SVM分类器来训练模型，以分辨目标和非目标。

4. 在测试图像上运行滑动窗口进行目标检测，将每个窗口的HOG特征向量输入到SVM分类器中，以确定窗口是否包含对象。

5. 对于每个被分类为目标的窗口，使用非极大抑制来抑制重叠的窗口，以得到最终的目标检测结果。

接下来，我们将按照上述流程编写代码，并讲解每个步骤的细节。

1. 收集和准备训练数据集

首先，我们需要准备包含目标和非目标数据集的文件夹。在这个案例中，我们使用了INRIA行人数据集，其中包含了正面和负面的图像样本。我们需要将这些图像分别放在两个文件夹中，并将它们的路径保存到.mat文件中。具体代码如下：

pos_folder = 'INRIAPerson/train_64x128_H96/pos';
neg_folder = 'INRIAPerson/train_64x128_H96/neg';

pos_img_files = dir(fullfile(pos_folder, '*.png'));
neg_img_files = dir(fullfile(neg_folder, '*.png'));

pos_img_paths = fullfile(pos_folder, {pos_img_files.name}');
neg_img_paths = fullfile(neg_folder, {neg_img_files.name}');

save('img_paths.mat', 'pos_img_paths', 'neg_img_paths');

2. 提取HOG特征

接下来，我们需要使用MATLAB内置的HOG函数来提取图像的HOG特征。对于每个图像，我们将其转换为灰度图像，并将其缩放为统一的大小。然后，我们使用HOG函数来提取特征向量，并将其保存到.mat文件中。具体代码如下：

% HOG parameters
cell_size = 8;
block_size = 2;
num_bins = 9;

% Load image paths
load('img_paths.mat');

% Extract HOG features for positive samples
pos_features = cell(numel(pos_img_paths), 1);
for i = 1:numel(pos_img_paths)
    img = imread(pos_img_paths{i});
    img_gray = rgb2gray(img);
    img_resized = imresize(img_gray, [128 64]);
    pos_features{i} = extractHOGFeatures(img_resized, 'CellSize', [cell_size cell_size],...
        'BlockSize', [block_size block_size], 'NumBins', num_bins);
end

% Extract HOG features for negative samples
neg_features = cell(numel(neg_img_paths), 1);
for i = 1:numel(neg_img_paths)
    img = imread(neg_img_paths{i});
    img_gray = rgb2gray(img);
    img_resized = imresize(img_gray, [128 64]);
    neg_features{i} = extractHOGFeatures(img_resized, 'CellSize', [cell_size cell_size],...
        'BlockSize', [block_size block_size], 'NumBins', num_bins);
end

% Save features to file
save('features.mat', 'pos_features', 'neg_features');

3. 使用SVM分类器来训练模型

现在，我们已经提取了所有图像的HOG特征，接下来我们需要使用SVM分类器来训练模型。在这个案例中，我们使用MATLAB内置的fitcsvm函数来训练线性SVM分类器。具体代码如下：

% Load features
load('features.mat');

% Combine positive and negative features
X = [vertcat(pos_features{:}); vertcat(neg_features{:})];
Y = [ones(numel(pos_features), 1); zeros(numel(neg_features), 1)];

% Train SVM classifier
svm = fitcsvm(X, Y, 'KernelFunction', 'linear');

% Save SVM model to file
save('svm_model.mat', 'svm');

4. 在测试图像上运行滑动窗口进行目标检测

现在，我们已经训练了SVM分类器，接下来我们需要在测试图像上运行滑动窗口进行目标检测。对于每个窗口，我们将其转换为灰度图像，并将其缩放为统一的大小。然后，我们使用HOG函数提取特征向量，并将其输入到SVM分类器中进行分类。如果窗口被分类为目标，我们将其添加到目标集合中。最后，我们使用非极大抑制算法来抑制重叠的窗口，以得到最终的目标检测结果。具体代码如下：

% Load SVM model
load('svm_model.mat');

% Load test image
test_img = imread('test.png');

% Sliding window parameters
window_size = [128 64];
step_size = [8 8];

% Initialize variables
target_boxes = [];
target_scores = [];

% Sliding window loop
for r = 1:step_size(1):(size(test_img, 1)-window_size(1))
    for c = 1:step_size(2):(size(test_img, 2)-window_size(2))
        
        % Extract window
        window = test_img(r:r+window_size(1)-1, c:c+window_size(2)-1, :);
        
        % Convert to grayscale and resize
        window_gray = rgb2gray(window);
        window_resized = imresize(window_gray, [128 64]);
        
        % Extract HOG features
        features = extractHOGFeatures(window_resized, 'CellSize', [cell_size cell_size],...
            'BlockSize', [block_size block_size], 'NumBins', num_bins);
        
        % Classify window using SVM
        score = predict(svm, features);
        
        % Add to targets if score is positive
        if score > 0
            target_boxes = [target_boxes; [c r window_size(2) window_size(1)]];
            target_scores = [target_scores; score];
        end
        
    end
end

% Apply non-maximum suppression
picked_boxes = selectStrongestBbox(target_boxes, target_scores, 'OverlapThreshold', 0.5);

% Draw bounding boxes on test image
figure;
imshow(test_img);
hold on;
for i = 1:size(picked_boxes, 1)
    rectangle('Position', picked_boxes(i, :), 'EdgeColor', 'r', 'LineWidth', 2);
end

5. 完整代码

最后，让我们将所有代码组合在一起，以获得完整的HOG+SVM目标检测代码。请注意，您需要提前下载INRIA行人数据集，并将其解压到正确的文件夹中。

% HOG+SVM Object Detection Example
% By Qishen Huang

%% 1. Collect and prepare training data

% Set folder paths
pos_folder = 'INRIAPerson/train_64x128_H96/pos';
neg_folder = 'INRIAPerson/train_64x128_H96/neg';

% Load positive and negative image paths
pos_img_files = dir(fullfile(pos_folder, '*.png'));
neg_img_files = dir(fullfile(neg_folder, '*.png'));
pos_img_paths = fullfile(pos_folder, {pos_img_files.name}');
neg_img_paths = fullfile(neg_folder, {neg_img_files.name}');
save('img_paths.mat', 'pos_img_paths', 'neg_img_paths');

%% 2. Extract HOG features

% Set HOG parameters
cell_size = 8;
block_size = 2;
num_bins = 9;

% Load image paths
load('img_paths.mat');

% Extract HOG features for positive samples
pos_features = cell(numel(pos_img_paths), 1);
for i = 1:numel(pos_img_paths)
    img = imread(pos_img_paths{i});
    img_gray = rgb2gray(img);
    img_resized = imresize(img_gray, [128 64]);
    pos_features{i} = extractHOGFeatures(img_resized, 'CellSize', [cell_size cell_size],...
        'BlockSize', [block_size block_size], 'NumBins', num_bins);
end

% Extract HOG features for negative samples
neg_features = cell(numel(neg_img_paths), 1);
for i = 1:numel(neg_img_paths)
    img = imread(neg_img_paths{i});
    img_gray = rgb2gray(img);
    img_resized = imresize(img_gray, [128 64]);
    neg_features{i} = extractHOGFeatures(img_resized, 'CellSize', [cell_size cell_size],...
        'BlockSize', [block_size block_size], 'NumBins', num_bins);
end

% Save features to file
save('features.mat', 'pos_features', 'neg_features');

%% 3. Train SVM classifier

% Load features
load('features.mat');

% Combine positive and negative features
X = [vertcat(pos_features{:}); vertcat(neg_features{:})];
Y = [ones(numel(pos_features), 1); zeros(numel(neg_features), 1)];

% Train SVM classifier
svm = fitcsvm(X, Y, 'KernelFunction', 'linear');

% Save SVM model to file
save('svm_model.mat', 'svm');

%% 4. Run sliding window on test image

% Load SVM model
load('svm_model.mat');

% Load test image
test_img = imread('test.png');

% Sliding window parameters
window_size = [128 64];
step_size = [8 8];

% Initialize variables
target_boxes = [];
target_scores = [];

% Sliding window loop
for r = 1:step_size(1):(size(test_img, 1)-window_size(1))
    for c = 1:step_size(2):(size(test_img, 2)-window_size(2))
        
        % Extract window
        window = test_img(r:r+window_size(1)-1, c:c+window_size(2)-1, :);
        
        % Convert to grayscale and resize
        window_gray = rgb2gray(window);
        window_resized = imresize(window_gray, [128 64]);
        
        % Extract HOG features
        features = extractHOGFeatures(window_resized, 'CellSize', [cell_size cell_size],...
            'BlockSize', [block_size block_size], 'NumBins', num_bins);
        
        % Classify window using SVM
        score = predict(svm, features);
        
        % Add to targets if score is positive
        if score > 0
            target_boxes = [target_boxes; [c r window_size(2) window_size(1)]];
            target_scores = [target_scores; score];
        end
        
    end
end

% Apply non-maximum suppression
picked_boxes = selectStrongestBbox(target_boxes, target_scores, 'OverlapThreshold', 0.5);

% Draw bounding boxes on test image
figure;
imshow(test_img);
hold on;
for i = 1:size(picked_boxes, 1)
    rectangle('Position', picked_boxes(i, :), 'EdgeColor', 'r', 'LineWidth', 2);
end

这就是一个完整的基于MATLAB的HOG+SVM目标检测案例，包含完整的代码及代码讲解。