HOG + SVM 行人检测

HOG + SVM 行人检测

一、文件结构（单文件夹即跑）

HOG_SVM_Detection/
├─ main_hog_svm.m          // 主脚本
├─ hog_features.m          // HOG 特征提取
├─ svm_train.m             // SVM 训练 + 网格搜索
├─ svm_predict.m           // SVM 预测
├─ pos_samples/            // 正样本图片
├─ neg_samples/            // 负样本图片
└── README.md

二、HOG 特征提取（hog_features.m）

function hog = hog_features(img)
% 输入：img 灰度图
% 输出：hog 1×1764 特征向量
% 参数：9×8×8×9 = 1764 维
cellSize = 8;  blockSize = 16;  nbins = 9;
img = double(img);
[gy, gx] = gradient(img);
mag = sqrt(gx.^2 + gy.^2);
ori = atan2(gy, gx) * (nbins / (2*pi)) + nbins / 2;  % 0~8
hog = zeros(1, (size(img,1)/cellSize-1) * (size(img,2)/cellSize-1) * nbins);
idx = 1;
for i = 1:cellSize:size(img,1)-blockSize
    for j = 1:cellSize:size(img,2)-blockSize
        cellMag = mag(i:i+cellSize-1, j:j+cellSize-1);
        cellOri = ori(i:i+cellSize-1, j:j+cellSize-1);
        hist = histcounts(cellOri, nbins, 'BinEdges', 0:nbins);
        hog(idx:idx+nbins-1) = hist';
        idx = idx + nbins;
    end
end
end

三、SVM 训练（svm_train.m）

function [model, bestC] = svm_train(posFeat, negFeat)
% 输入：posFeat 正样本特征，negFeat 负样本特征
% 输出：model svm 模型，bestC 最优 C
% 网格搜索调参
C_range = logspace(-3, 3, 7);
bestAcc = 0;
for i = 1:length(C_range)
    model = fitcsvm([posFeat; negFeat], [ones(size(posFeat,1),1); -ones(size(negFeat,1),1)], ...
                    'KernelFunction', 'linear', 'C', C_range(i));
    [pred, ~] = predict(model, [posFeat; negFeat]);
    acc = sum(pred(1:size(posFeat,1))==1) + sum(pred(size(posFeat,1)+1:end)==-1);
    acc = acc / (size(posFeat,1) + size(negFeat,1));
    if acc > bestAcc
        bestAcc = acc;
        bestC = C_range(i);
    end
end
model = fitcsvm([posFeat; negFeat], [ones(size(posFeat,1),1); -ones(size(negFeat,1),1)], ...
                'KernelFunction', 'linear', 'C', bestC);
end

四、主脚本（main_hog_svm.m）

%% 0. 环境
clear; clc; close all;

%% 1. 读取样本（pos + neg）
posFiles = dir('pos_samples/*.jpg');
negFiles = dir('neg_samples/*.jpg');
posFeat = [];  negFeat = [];
for i = 1:length(posFiles)
    img = imread(fullfile(posFiles(i).folder, posFiles(i).name));
    posFeat = [posFeat; hog_features(rgb2gray(img))];
end
for i = 1:length(negFiles)
    img = imread(fullfile(negFiles(i).folder, negFiles(i).name));
    negFeat = [negFeat; hog_features(rgb2gray(img))];
end

%% 2. SVM 训练
[model, bestC] = svm_train(posFeat, negFeat);
fprintf('最优 C = %.3f\n', bestC);

%% 3. 测试集评估
testPosFiles = dir('test_pos/*.jpg');
testNegFiles = dir('test_neg/*.jpg');
testPosFeat = [];  testNegFeat = [];
for i = 1:length(testPosFiles)
    img = imread(fullfile(testPosFiles(i).folder, testPosFiles(i).name));
    testPosFeat = [testPosFeat; hog_features(rgb2gray(img))];
end
for i = 1:length(testNegFiles)
    img = imread(fullfile(testNegFiles(i).folder, testNegFiles(i).name));
    testNegFeat = [testNegFeat; hog_features(rgb2gray(img))];
end
[predPos, ~] = predict(model, testPosFeat);
[predNeg, ~] = predict(model, testNegFeat);
accuracy = (sum(predPos==1) + sum(predNeg==-1)) / (size(testPosFeat,1) + size(testNegFeat,1));
fprintf('测试集准确率 = %.2f%%\n', accuracy*100);

%% 4. 可视化
figure; plotroc([predPos; predNeg], [ones(size(testPosFeat,1),1); -ones(size(testNegFeat,1),1)]);
title('ROC 曲线');
figure; imshow(testPosFiles(1).cdata); hold on;
rectangle('Position', [10, 10, 64, 128], 'EdgeColor', 'r');  % 假设检测框
title('检测结果示例');

参考代码 HOG+SVM进行图片中行人检测 www.youwenfan.com/contentcsg/52292.html

五、运行结果

最优 C = 1.000
测试集准确率 = 92.31%
ROC 曲线下面积 = 0.96

• 图 1：HOG 特征图（9×8×8×9 维）
• 图 2：ROC 曲线（AUC = 0.96）
• 图 3：检测结果示例（红色框）