【opencv-python】hog+svm实现行人识别

import cv2
import os
import numpy as np
import hog_discriptor


def get_path():
    print('loading data...')
    # 正样本文件夹路径
    PosImgPath = "D:\\A\\python\\opencv\\tests\\hog_pedestran_detect_python\\Positive"
    # 负样本
    NegImgPath = "D:\\A\\python\\opencv\\tests\\hog_pedestran_detect_python\\Negative"
    # 文件夹所有图片文件名
    PosImgList = os.listdir(PosImgPath)
    NegImgList = os.listdir(NegImgPath)
    print('Number of positive samples:', len(PosImgList))
    print('Number of negative samples:', len(NegImgList))
    # 合并数据路径 添加对应标签
    samplePath = []
    labels = []
    for f in PosImgList:
        samplePath.append('Positive\\' + f)
        labels.append(1)
    for f in NegImgList:
        samplePath.append('Negative\\' + f)
        labels.append(-1)
    labels = np.array(labels)

    return samplePath, labels


def extract_hog(samplePath):
    print('extracting hog...')
    # hog参数
    winSize = (64, 128)     # 窗口大小
    blockSize = (16, 16)    # 块大小
    blockStride = (8, 8)    # 块滑动增量
    cellSize = (8, 8)       # 胞元大小
    Bin = 9                 # 梯度方向数
    # 创建hog
    hogDescriptor = cv2.HOGDescriptor(winSize, blockSize, blockStride, cellSize, Bin)
    # 获取正负样本hog特征
    hogFeature = []
    for f in samplePath:
        img = cv2.imread(f, cv2.IMREAD_GRAYSCALE)
        img = cv2.resize(img, (64, 128))
        discriptor = hogDescriptor.compute(img)
        hogFeature.append(discriptor)
    # 转为numpy数组
    hogFeature = np.array(hogFeature)

    return hogFeature


def train_svm(hogFeature, labels):
    # 创建svm
    svm = cv2.ml.SVM_create()
    svm.setKernel(cv2.ml.SVM_LINEAR)    # 使用线性核
    svm.setP(0.01)       # 练习集中的特征向量和拟合出来的超平面的间隔须要小于p
    svm.setC(0.01)      # 异常值处罚因子 一般取10的n次幂 default=1.0 越接近0惩罚越大
    svm.setType(cv2.ml.SVM_EPS_SVR)     # SVM类型 SVR支持向量回归机
    # 训练
    print('training...')
    svm.train(hogFeature, cv2.ml.ROW_SAMPLE, labels)
    print('done.')

    return svm


def get_svm_detecter(svm):
    # 获取支持向量
    SupportVector = svm.getSupportVectors()
    SupportVector = np.transpose(SupportVector)
    # rho-决策函数常数项b的相反数
    rho, alpha, _ = svm.getDecisionFunction(0)
    # 返回分类器
    return np.append(alpha * SupportVector, [[-rho]], 0)


def detect(MyDecter):
    imageSrc = cv2.imread('TestData\\000001.jpg', cv2.IMREAD_COLOR)

    myHog = cv2.HOGDescriptor()
    myHog.setSVMDetector(MyDecter)
    # 设置步长和padding、每次的缩放比例
    objects, _ = myHog.detectMultiScale(imageSrc, winStride=(8, 8), padding=(16, 16), scale=1.04)
    for (x, y, w, h) in objects:
        cv2.rectangle(imageSrc, (x, y), (x+w, y+h), (255, 0, 0), 2)

    cv2.imshow('img', imageSrc)
    cv2.waitKey(0)


if __name__ == '__main__':
    samplePath, labels = get_path()
    hogFeature = extract_hog(samplePath)
    svm = train_svm(hogFeature, labels)
    MyDecter = get_svm_detecter(svm)
    detect(MyDecter)