摔倒检测+yolov5

数据集来源

开始使用目标检测检测人体，再通过宽高比值变化来判断是否摔倒，该方法在密集场景不管用，所以想着找一个摔倒数据集来进行检测是否摔倒，附上数据集来源：
中国华录杯——定向算法赛（人体摔倒姿态识别）这个训练集数据集我没有找到，官网下载的也只是未标注的测试集
Le2i Fall detection Dataset
• Annotations

• 标签格式为txt格式
• 标签前两行为跌倒帧数的起始与终止点，也存在没有这两行的文件
• 之后一行就是六个字段，帧编号， XX，人物bbx

Le2i Fall detection Dataset 原文下载速度不敢恭维，附上 百度云下载链接, 提取码：qghr
中国华录杯——定向算法赛（人体摔倒姿态识别）测试集，无标注，自己找的一些算法打的标注，还算过得去，需要自取百度云 百度云下载链接 , 提取码：asdh
以上两个数据集要是侵权了会删除

数据的处理

对Le2i Fall detection Dataset数据转换成voc格式的xml文件：

1. 数据集解压后如下图，不知道有两个场景是没有标注文件的，号称假摔，所以我把他删了
   
   
2. 解压后运行如下代码

# coding=utf-8
"""
@Time: 2021/3/9 10:23
@IDE: PyCharm
@Author: chengzhang
@File: processVideo.py
"""
import os
from tqdm import tqdm
import cv2
from xml.dom.minidom import Document

count = 0


def process(fileDir):
    dataset = [os.path.join(fileDir, i) for i in os.listdir(fileDir)]

    for dir in dataset:
        Annotation = os.path.join(dir, "Annotation_files")
        Annotations = [os.path.join(Annotation, i) for i in os.listdir(Annotation)]
        for ann in tqdm(Annotations):
            txt = open(ann).readlines()
            if len(txt[0].split(',')) < 2:
                videoPath = ann.replace('Annotation_files', 'Videos').replace('.txt', '.avi')
                start, end = int(txt[0].replace('\n', '')), int(txt[1].replace('\n', ''))
                video(videoPath, txt, start, end)


def video(videoPath, txt, start, end):
    global count
    vs = cv2.VideoCapture(videoPath)
    ind = 1
    while True:
        success, frame = vs.read()
        if frame is None:
            break
        if (ind + 8) >= start and ind <= end:
            Jpg = '{:06d}.jpg'.format(count)
            outPutJpg = os.path.join(save, Jpg)
            outPutXML = outPutJpg.replace('jpg', 'xml')
            cv2.imwrite(outPutJpg, frame)
            count += 1
            imgShape = frame.shape
            info = txt[ind].replace('\n', '').split(',')
            _, _, xmin, ymin, xmax, ymax = info
            bboxsInfo = [[int(xmin), int(ymin), int(xmax), int(ymax)]]
            bbox2xml(bboxsInfo, imgShape, Jpg, outPutXML)
        ind += 1

def bbox2xml(bboxsInfo, imgShape, outPutJpg, outPutXML):
    # for boxInfo in bboxsInfo:
    xmlBuilder = Document()
    annotation = xmlBuilder.createElement("annotation")  # 创建annotation标签
    xmlBuilder.appendChild(annotation)
    Pheight, Pwidth, Pdepth = imgShape

    folder = xmlBuilder.createElement("folder")  # folder标签
    folderContent = xmlBuilder.createTextNode("VOC2007")
    folder.appendChild(folderContent)
    annotation.appendChild(folder)

    filename = xmlBuilder.createElement("filename")  # filename标签
    filenameContent = xmlBuilder.createTextNode(outPutJpg)
    filename.appendChild(filenameContent)
    annotation.appendChild(filename)

    size = xmlBuilder.createElement("size")  # size标签
    width = xmlBuilder.createElement("width")  # size子标签width
    widthContent = xmlBuilder.createTextNode(str(Pwidth))
    width.appendChild(widthContent)
    size.appendChild(width)
    height = xmlBuilder.createElement("height")  # size子标签height
    heightContent = xmlBuilder.createTextNode(str(Pheight))
    height.appendChild(heightContent)
    size.appendChild(height)
    depth = xmlBuilder.createElement("depth")  # size子标签depth
    depthContent = xmlBuilder.createTextNode(str(Pdepth))
    depth.appendChild(depthContent)
    size.appendChild(depth)
    annotation.appendChild(size)

    for bbox in bboxsInfo:
        label = "fall"
        box_xmin, box_ymin, box_xmax, box_ymax = bbox

        object = xmlBuilder.createElement("object")
        picname = xmlBuilder.createElement("name")
        nameContent = xmlBuilder.createTextNode(label)
        picname.appendChild(nameContent)
        object.appendChild(picname)
        pose = xmlBuilder.createElement("pose")
        poseContent = xmlBuilder.createTextNode("Unspecified")
        pose.appendChild(poseContent)
        object.appendChild(pose)
        truncated = xmlBuilder.createElement("truncated")
        truncatedContent = xmlBuilder.createTextNode("0")
        truncated.appendChild(truncatedContent)
        object.appendChild(truncated)
        difficult = xmlBuilder.createElement("difficult")
        difficultContent = xmlBuilder.createTextNode("0")
        difficult.appendChild(difficultContent)
        object.appendChild(difficult)

        bndbox = xmlBuilder.createElement("bndbox")
        xmin = xmlBuilder.createElement("xmin")
        xminContent = xmlBuilder.createTextNode(str(box_xmin))
        xmin.appendChild(xminContent)
        bndbox.appendChild(xmin)
        ymin = xmlBuilder.createElement("ymin")
        yminContent = xmlBuilder.createTextNode(str(box_ymin))
        ymin.appendChild(yminContent)
        bndbox.appendChild(ymin)
        xmax = xmlBuilder.createElement("xmax")
        xmaxContent = xmlBuilder.createTextNode(str(box_xmax))
        xmax.appendChild(xmaxContent)
        bndbox.appendChild(xmax)
        ymax = xmlBuilder.createElement("ymax")
        ymaxContent = xmlBuilder.createTextNode(str(box_ymax))
        ymax.appendChild(ymaxContent)
        bndbox.appendChild(ymax)
        object.appendChild(bndbox)

        annotation.appendChild(object)
    f = open(outPutXML, 'w')
    xmlBuilder.writexml(f, indent='\t', newl='\n', addindent='\t', encoding='utf-8')

    f.close()


if __name__ == '__main__':
    fileDir = r"E:\数据集\FallDataset\dataset"  # 源数据集路径
    save = r"E:\数据集\FallDataset\FallDataset"  # 保存voc格式路径
    process(fileDir)

模型训练

1. 使用U神的yolov5，链接
2. 将voc格式的数据集转换成yolov5所需要的格式，数据格式如下：其中txt标注为下面代码生成的归一化数据
   

# coding=utf-8
"""
@Time: 2020/9/17 9:07
@IDE: PyCharm
@Author: chengzhang
@File: voc2v5.py
"""
import os
import glob
from tqdm import tqdm
import xml.etree.ElementTree as ET
import shutil
import argparse
from random import shuffle
import cv2


def check_x_y_w_h(norm_x, norm_y, norm_w, norm_h):
    norm_x = 0.00001 if norm_x <= 0 else norm_x
    norm_x = 0.99999 if norm_x >= 1 else norm_x
    norm_y = 0.00001 if norm_y <= 0 else norm_y
    norm_y = 0.99999 if norm_y >= 1 else norm_y
    norm_w = 0.00001 if norm_w <= 0 else norm_w
    norm_w = 0.99999 if norm_w >= 1 else norm_w
    norm_h = 0.00001 if norm_h <= 0 else norm_h
    norm_h = 0.99999 if norm_h >= 1 else norm_h
    return norm_x, norm_y, norm_w, norm_h


def convert(size, box):
    dw = 1. / (size[0])
    dh = 1. / (size[1])
    x = (box[0] + box[1]) / 2.0 - 1
    y = (box[2] + box[3]) / 2.0 - 1
    w = box[1] - box[0]
    h = box[3] - box[2]
    x = x * dw
    w = w * dw
    y = y * dh
    h = h * dh
    x, y, w, h = check_x_y_w_h(x, y, w, h)
    return (x, y, w, h)


def write_txt(train_xml, dataset):
    with open(dataset, "w", encoding="utf-8") as f:
        for xml in tqdm(train_xml):

            f.write(xml.replace('.xml', '.jpg') + '\n')
            target_label_txt = xml.replace('.xml', '.txt')
            img = cv2.imread(xml.replace('.xml', '.jpg'))
            h, w, c = img.shape
            # 输入文件xml
            in_file = open(xml)
            # 输出label txt
            out_file = open(target_label_txt, 'w')
            tree = ET.parse(in_file)
            root = tree.getroot()
            for obj in root.iter('object'):
                try:
                    cls = obj.find('name').text
                except:
                    continue

                if cls not in classes:
                    continue
                xmlbox = obj.find('bndbox')
                b = (float(xmlbox.find('xmin').text), float(xmlbox.find('xmax').text), float(xmlbox.find('ymin').text),
                     float(xmlbox.find('ymax').text))
                bb = convert((w, h), b)
                out_file.write(str(0) + " " + " ".join([str(a) for a in bb]) + '\n')


def check_dir(dir):
    try:
        shutil.rmtree(dir)
    except:
        os.makedirs(dir)
    if not os.path.exists(dir):
        os.makedirs(dir)


def parse_annotation(img_dir):
    all_imgs = []
    all_xmls = []
    for tmpdir in img_dir:
        for home, dirs, files in os.walk(tmpdir):
            for filename in files:
                # 文件名列表，包含完整路径
                if filename.endswith((".jpg", ".jpeg", ".png")):
                    filePath = os.path.join(home, filename)
                    all_imgs.append(filePath)
                elif filename.endswith(".xml"):
                    filePath = os.path.join(home, filename)
                    all_xmls.append(filePath)

    return all_xmls


if __name__ == '__main__':
    classes = ['fall']

    parser = argparse.ArgumentParser()
    parser.add_argument("--train_txt", default="/data/share/imageAlgorithm/zhangcheng/code/yolov5/data/train_fall.txt")
    parser.add_argument("--test_txt", default="/data/share/imageAlgorithm/zhangcheng/code/yolov5/data/val_fall.txt")

    flags = parser.parse_args()

    imgPathList = [
        "/data/share/imageAlgorithm/zhangcheng/dataset/falldataset/FallDataset/",
        "/data/share/imageAlgorithm/zhangcheng/dataset/falldataset/FallDataset_A/"
    ]

    dataset = parse_annotation(imgPathList)
    shuffle(dataset)
    num = len(dataset)
    trainNum = int(0.9 * num)

    trainList = dataset[: trainNum]
    testList = dataset[trainNum:]

    write_txt(trainList, flags.train_txt)
    write_txt(testList, flags.test_txt)

3. 根据上面处理的数据进行训练

CUDA_VISIBLE_DEVICES=4,5,6,7 nohup python train.py --img 640 --batch 16 --epochs 600 --data ./data/fallDown.yaml --weights ./weights/yolov5l.pt > train_fall.log 2>&1 &

4. 训练曲线如下：
   30个epoch结果