基于MATLAB的HOG+SVM行人检测

基于MATLAB的HOG+SVM行人检测

%% 1. 数据准备与预处理
% 加载正负样本数据集（需自行准备）
load('pedestrian_dataset.mat'); % 包含trainData/trainLabels和testData/testLabels

% 图像预处理函数
function processedImg = preprocess(img)
    % 灰度化
    grayImg = rgb2gray(img);
    % 高斯滤波去噪
    blurredImg = imgaussfilt(grayImg, 2);
    % 直方图均衡化
    equalizedImg = histeq(blurredImg);
    processedImg = imresize(equalizedImg, [64, 128]); % 统一尺寸
end

% 预处理训练集
trainDataPreprocessed = arrayfun(@(i) preprocess(trainData(:,:,i)), 1:size(trainData,3), 'UniformOutput', false);
testDataPreprocessed = arrayfun(@(i) preprocess(testData(:,:,i)), 1:size(testData,3), 'UniformOutput', false);

%% 2. HOG特征提取
% 定义HOG参数
cellSize = [8,8];
blockSize = [2,2];
blockStride = [1,1];
numBins = 9;

% 提取训练特征
trainFeatures = [];
for i = 1:numel(trainDataPreprocessed)
    img = trainDataPreprocessed{i};
    hogFeat = extractHOGFeatures(img, 'CellSize', cellSize, ...
        'BlockSize', blockSize, 'BlockStride', blockStride, 'NumBins', numBins);
    trainFeatures = [trainFeatures; hogFeat];
end

% 提取测试特征
testFeatures = [];
for i = 1:numel(testDataPreprocessed)
    img = testDataPreprocessed{i};
    hogFeat = extractHOGFeatures(img, 'CellSize', cellSize, ...
        'BlockSize', blockSize, 'BlockStride', blockStride, 'NumBins', numBins);
    testFeatures = [testFeatures; hogFeat];
end

%% 3. SVM分类器训练
% 训练参数设置
svmOptions = statset('Display','iter');
svmModel = fitcsvm(trainFeatures, trainLabels, ...
    'KernelFunction', 'linear', ...
    'BoxConstraint', 1, ...
    'Standardize', true, ...
    'Options', svmOptions);

%% 4. 检测器构建
% 创建检测器对象
detector = vision.PeopleDetector('UprightPeople_96x48', 'MergeThreshold', 30);

% 或使用自定义训练的SVM检测器
detector = vision.ObjectDetector('Custom', 'Model', svmModel, ...
    'FeatureExtractor', @hogFeatureExtractor);

%% 5. 实时检测演示
% 打开摄像头
vid = videoinput('winvideo', 1, 'YUY2_640x480');
vid.FramesPerTrigger = Inf;
vid.ReturnedColorspace = 'rgb';
preview(vid); % 预览画面

% 创建显示窗口
h = figure('Name','Real-time Pedestrian Detection', ...
    'Position',[100 100 800 600]);

try
    while ishandle(h)
        % 获取视频帧
        frame = getsnapshot(vid);
        
        % 检测行人
        [bboxes, scores] = detect(detector, frame);
        
        % 筛选高置信度结果
        validIdx = scores > 0.4;
        bboxes = bboxes(validIdx,:);
        scores = scores(validIdx);
        
        % 绘制检测结果
        if ~isempty(bboxes)
            frame = insertObjectAnnotation(frame, 'rectangle',...
                bboxes, sprintf('Score %.2f', scores(1)),...
                'Color','red','LineWidth',3);
        end
        
        % 显示结果
        imshow(frame);
        title(sprintf('Detected %d pedestrians',size(bboxes,1)));
        drawnow;
    end
catch
end

%% 6. 清理资源
stop(vid);
delete(vid);
clear vid;

%% 辅助函数：HOG特征提取
function features = hogFeatureExtractor(img)
    % 图像预处理
    grayImg = rgb2gray(img);
    % 提取HOG特征
    [features, ~] = extractHOGFeatures(grayImg, 'CellSize',[8,8],...
        'BlockSize',[2,2], 'BlockStride',[1,1], 'NumBins',9);
end

---

关键参数说明（基于搜索结果优化）| 参数名 | 推荐值 | 作用说明 |

|----------------|-----------|------------------------------|

| cellSize| | 梯度方向统计单元尺寸 |

| blockSize| | 归一化块尺寸 |

| numBins| 9 | 梯度方向量化级数 |

| blockStride| | 块间滑动步长 |

| MergeThreshold| 30 | 检测框合并阈值 |

---

性能优化策略1. 多尺度检测（改进检测精度）：

detector = vision.PeopleDetector('UprightPeople_96x48', 'NumScales', 5, 'ScaleStep', 1.2);

1. GPU加速（需Parallel Computing Toolbox）：

gpuImg = gpuArray(frame);
[bboxes, scores] = detect(detector, gpuImg);

1. 非极大值抑制（减少重叠检测）：

% 自定义NMS函数
function pick = nms(boxes, overlapThresh)
    if isempty(boxes)
        pick = [];
        return;
    end
    x1 = boxes(:,1); y1 = boxes(:,2);
    x2 = boxes(:,3); y2 = boxes(:,4);
    areas = (x2-x1+1).*(y2-y1+1);
    [~, idx] = sort(scores, 'descend');
    pick = [];
    while ~isempty(idx)
        last = length(idx);
        i = idx(last);
        pick = [pick; i];
        xx1 = max(x1(i), x1(idx(1:last-1)));
        yy1 = max(y1(i), y1(idx(1:last-1)));
        xx2 = min(x2(i), x2(idx(1:last-1)));
        yy2 = min(y2(i), y2(idx(1:last-1)));
        w = max(0, xx2 - xx1 + 1);
        h = max(0, yy2 - yy1 + 1);
        overlap = (w.*h)./areas(idx(1:last-1));
        idx(idx(overlap > overlapThresh)) = [];
    end
end

典型应用场景优化建议1. 监控视频处理：

% 视频文件处理
video = VideoReader('traffic.mp4');
while hasFrame(video)
    frame = readFrame(video);
    % 降采样加速
    smallFrame = imresize(frame, 0.5);
    % 检测...
end

1. 移动端优化：

% 降低分辨率
inputSize = [160, 120];
detector = vision.PeopleDetector('UprightPeople_48x32', 'InputSize', inputSize);

---

参考代码 HOG 行人检测代码 www.youwenfan.com/contentcni/65187.html

常见问题解决方案1. 漏检问题：

• 调整检测尺度范围（NumScales参数）
• 增加训练样本多样性

1. 误检问题： 提高置信度阈值（MergeThreshold） 启用阴影去除算法：

frame = removeShadows(frame, 'Method', 'AdaptiveThreshold');

1. 实时性优化： 使用积分图像加速：

hogParams = struct('UseIntegralImage', true);
detector = vision.PeopleDetector('UprightPeople_96x48', 'HOGParams', hogParams);

---

扩展应用方向1. 深度学习融合：

% 使用预训练CNN提取特征
net = alexnet;
inputSize(net) = [227 227 3];
features = activations(net, imresize(frame, [227 227]), 'fc7', 'OutputAs', 'rows');

1. 多传感器融合：

% 结合雷达数据优化检测
radarData = readRadarData('sensor.mat');
[bboxes, scores] = detectWithSensorData(detector, frame, radarData);

1. 动态场景适应：

% 基于光流法的运动补偿
opticFlow = opticalFlowLK;
flow = estimateFlow(opticFlow, prevFrame);
compensatedFrame = applyFlow(frame, flow);