yolov5行人检测算法

PaddleDetection/pphuman_mot.md at release/2.6 · PaddlePaddle/PaddleDetection · GitHubObject Detection toolkit based on PaddlePaddle. It supports object detection, instance segmentation, multiple object tracking and real-time multi-person keypoint detection. - PaddleDetection/pphuman_mot.md at release/2.6 · PaddlePaddle/PaddleDetectionhttps://github.com/PaddlePaddle/PaddleDetection/blob/release/2.6/deploy/pipeline/docs/tutorials/pphuman_mot.mdGitHub - JialeCao001/PedSurvey: From Handcrafted to Deep Features for Pedestrian Detection: A Survey (TPAMI 2021)From Handcrafted to Deep Features for Pedestrian Detection: A Survey (TPAMI 2021) - GitHub - JialeCao001/PedSurvey: From Handcrafted to Deep Features for Pedestrian Detection: A Survey (TPAMI 2021)https://github.com/JialeCao001/PedSurvey行人检测综述_denghe1122的博客-CSDN博客PART Ifrom: http://www.cnblogs.com/molakejin/p/5708791.html行人检测具有极其广泛的应用：智能辅助驾驶，智能监控，行人分析以及智能机器人等领域。从2005年以来行人检测进入了一个快速的发展阶段，但是也存在很多问题还有待解决，主要还是在性能和速度方面还不能达到一个权衡。近年，以谷歌为首的自动驾驶技术的研发正如火如荼的进行，这也迫https://blog.csdn.net/denghecsdn/article/details/77987627基于FairMOT实现人流量统计 - 飞桨AI Studio本项目基于PaddleDetection FairMOT实现动态场景和静态场景下的人流量统计，提供从 “模型选择→模型优化→模型部署” 的全流程指导，模型可以直接或经过少量数据微调后用于相关任务。 - 飞桨AI Studiohttps://aistudio.baidu.com/aistudio/projectdetail/2421822?channelType=0&channel=0VOC-COCO-MOT20 - 飞桨AI StudioVOC-COCO-MOT20 - 飞桨AI Studiohttps://aistudio.baidu.com/aistudio/datasetdetail/47128

时耕科技-中国领先的商业数智化服务提供商 - 时耕科技时耕科技-中国领先的商业数智化服务提供商https://www.timework.cn/col.jsp?id=157YOLOv5训练自己的数据集(超详细)_yolo数据集_AI追随者的博客-CSDN博客一、准备深度学习环境本人的笔记本电脑系统是：Windows10首先进入YOLOv5开源网址，手动下载zip或是git clone 远程仓库，本人下载的是YOLOv5的5.0版本代码，代码文件夹中会有requirements.txt文件，里面描述了所需要的安装包。本文最终安装的pytorch版本是1.8.1，torchvision版本是0.9.1，python是3.7.10，其他的依赖库按照requirements.txt文件安装即可。...https://blog.csdn.net/qq_40716944/article/details/118188085

yolov5 训练crowded human 【visible body detection】_yolov5训练crowdh_CV-杨帆的博客-CSDN博客使用yolov5训练crowded human中的head与visible bodyhttps://blog.csdn.net/WhiffeYF/article/details/124502681本项目来源于明厨亮灶这样一个背景的视频监控项目，对厨房中是否存在有人场景做一个识别，打算使用yolov5算法。

1.数据集

目前行人检测的数据集包括coco_person，voc_person，mot20/16、17，CrowdHuman ， HIEVE，Caltech Pedestrian，CityPersons， CHUK-SYSU，PRW，ETHZ。

2.数据处理

采用coco-voc-mot20数据集，一共是41856张图，其中训练数据37736张图，验证数据3282张图，测试数据838张。

2.1 将上述数据中的train/val/test转成包含图片的txt。

# coding:utf-8

import os
import random
import argparse

parser = argparse.ArgumentParser()
# xml文件的地址，根据自己的数据进行修改 xml一般存放在Annotations下
parser.add_argument('--xml_path', default='/home/imcs/local_disk/D0011/Annotations', type=str, help='input xml label path')
# 数据集的划分，地址选择自己数据下的ImageSets/Main
parser.add_argument('--txt_path', default='dataSet', type=str, help='output txt label path')
opt = parser.parse_args([])

trainval_percent = 0.98 # 剩下的0.02就是测试集
train_percent = 0.9
xmlfilepath = opt.xml_path
txtsavepath = opt.txt_path
total_xml = os.listdir(xmlfilepath)
if not os.path.exists(txtsavepath):
    os.makedirs(txtsavepath)

num = len(total_xml)
list_index = range(num)
tv = int(num * trainval_percent)
tr = int(tv * train_percent)
trainval = random.sample(list_index, tv)
train = random.sample(trainval, tr)

file_trainval = open(txtsavepath + '/trainval.txt', 'w')
file_test = open(txtsavepath + '/test.txt', 'w')
file_train = open(txtsavepath + '/train.txt', 'w')
file_val = open(txtsavepath + '/val.txt', 'w')

for i in tqdm(list_index):
#     import pdb;pdb.set_trace()
    name = Path(total_xml[i]).stem + '\n'
    
    if i in trainval:
        file_trainval.write(name)
        if i in train:
            file_train.write(name)
        else:
            file_val.write(name)
    else:
        file_test.write(name)

file_trainval.close()
file_train.close()
file_val.close()
file_test.close()

2.2 将xml数据转成yolo数据格式

sets = ['train', 'val', 'test']
classes = ["person"]   # 改成自己的类别
abs_path = os.getcwd()
print(abs_path)
 
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
    return x, y, w, h
 
def convert_annotation(image_id):
    in_file = open(abs_path+'/Annotations/%s.xml' % (image_id), encoding='UTF-8')
    out_file = open(abs_path+'/Yolo/%s.txt' % (image_id), 'w')
    tree = ET.parse(in_file)
    root = tree.getroot()
    size = root.find('size')
    w = int(size.find('width').text)
    h = int(size.find('height').text)
    for obj in root.iter('object'):
        # difficult = obj.find('difficult').text
        try:
            difficult = obj.find('Difficult').text
        except:
            difficult = 0
        cls = obj.find('name').text
        if cls not in classes or int(difficult) == 1:
            continue
        cls_id = classes.index(cls)
        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))
        b1, b2, b3, b4 = b
        # 标注越界修正
        if b2 > w:
            b2 = w
        if b4 > h:
            b4 = h
        b = (b1, b2, b3, b4)
        bb = convert((w, h), b)
        out_file.write(str(cls_id) + " " + " ".join([str(a) for a in bb]) + '\n')

        
wd = getcwd()
for image_set in tqdm(sets):
    image_ids = [line.strip() for line in  open('./dataSet/%s.txt' % (image_set)).readlines()]
    list_file = open('%s_yolo.txt' % (image_set), 'w')
    for image_id in image_ids:
        try:
            list_file.write(abs_path + '/JPEGImages/%s.jpg\n' % (image_id))

            convert_annotation(image_id)
        except:
            continue
    list_file.close()

2.3 更改数据集名称

../datasets/coco128/images/im0.jpg  # image
../datasets/coco128/labels/im0.txt  # label

yolo会自动的把image换成labels，去找标签名，因此数据集名称要换成labels和images。

2.5 crownhuman数据

使用YOLOv5-Tools中的CrowHuman2YOLO中代码

bash ./prepare_data.sh 608x608

生成了crownhuman-608x608，里面已经是yolo格式了，但是添加到上面的数据集里，还需要一些操作。这其中图片数据：19370，文本数据：19370，其中test数据：4370，train数据：15000.

2.5.1 将image和txt分别提取

import os
import shutil

# 源文件夹路径
source_folder = "/home/imcs/local_disk/crownhuman/crowdhuman-608x608/"
# 目标文件夹路径
target_folder = "/home/imcs/local_disk/crownhuman/images"

# 遍历源文件夹中的所有文件
for filename in os.listdir(source_folder):
    file_path = os.path.join(source_folder, filename)
    # 如果这是一张图片，则迁移该图片到目标文件夹
    if os.path.isfile(file_path) and filename.lower().endswith(('.jpg', '.jpeg', '.png', '.gif')):
        shutil.move(file_path, target_folder)
        print('Moved image:', file_path)


import os
import shutil

# 源文件夹路径
source_folder = "/home/imcs/local_disk/crownhuman/crowdhuman-608x608/"
# 目标文件夹路径
target_folder = "/home/imcs/local_disk/crownhuman/labels"

# 遍历源文件夹中的所有文件
for filename in os.listdir(source_folder):
    file_path = os.path.join(source_folder, filename)
    # 如果这是一个txt文件，则迁移该文件到目标文件夹
    if os.path.isfile(file_path) and filename.lower().endswith('.txt'):
        shutil.move(file_path, target_folder)
#         print('Moved txt file:', file_path)

2.5.2 对路径进行更新

import os

# 源 txt 文件路径
source_file = "/home/imcs/local_disk/crownhuman/train.txt"
# 目标 txt 文件路径
target_file = "/home/imcs/local_disk/crownhuman/train_yolo.txt"

# 图片路径前缀（例如：原图路径为 "/path/to/image.jpg"，更新后的路径为 "/new_path/to/image.jpg"）
prefix = "/home/imcs/local_disk/crownhuman/images"

# 打开源 txt 文件和目标 txt 文件
with open(source_file, 'r') as f, open(target_file, 'w') as t:
#     import pdb;pdb.set_trace()
    # 逐行读取源 txt 文件中的内容
    for line in f:
        # 删除每行末尾的换行符
        line = line.strip()
        # 如果该行内容是一个图片路径，则更新该路径并写入目标 txt 文件
        if line.lower().endswith(('.jpg', '.jpeg', '.png', '.gif')):
            new_line = os.path.join(prefix, os.path.basename(line))
            t.write(new_line + '\n')
            print('Updated image path:', line, '->', new_line)
        else:
            t.write(line + '\n')

2.5.3 对labels中的0标签进行修改，0 表示head，1表示person

import os

input_dir_path = '/home/imcs/local_disk/crownhuman/labels_all'
output_dir_path = '/home/imcs/local_disk/crownhuman/labels'

for filename in os.listdir(input_dir_path):
    if not filename.endswith('.txt'):
        continue
        
    input_file_path = os.path.join(input_dir_path, filename)
#     output_file_path = os.path.join(output_dir_path, filename.replace('.txt', '_processed.txt'))
    output_file_path = os.path.join(output_dir_path, filename)
    
    with open(input_file_path) as input_file, open(output_file_path, 'w') as output_file:
        for line in input_file:
            row = line.strip().split()
            label = int(row[0])
            if label == 1:
                output_file.write(' '.join([row[0].replace("1","0")]+row[1:]) + '\n')
print('Done!')

3.环境安装

python3.7 cuda10.1  

pip install -r requirements

torch==1.8.1_cu101

torchvision==0.9.1_cu101

tqdm==4.64.0

thop==0.1.1.post2207130030

matplotlib==3.2.2

numpy==1.18.5

opencv-python==4.1.1

pillow==7.1.2

pyyaml==5.3.1

requests==2.23.0

scipy==1.4.1

tqdm==4.64.0

tensorboard==2.4.1

pandas==1.1.4

seaborn==0.11.0

升级glic/libstdc++.so.6

Downloading https://ultralytics.com/assets/Arial.ttf to /home/imcs/.config/Ultralytics/Arial.ttf

4.模型

 一般考虑在端侧部署的话就用yolov5s，和大模型不少于1b的参数量相比，检测模型的参数量缺少很少。

yolov5代码写的还是很好的，到目前为止，能在k80上分布式DDP无痛跑起来的代码可不多，mm系列都需要稍微改一下。

main->
resume->torch.load(last,map_location='cpu')
DDP->
device=select_device(device)->
dist=init_process_group(backend='nccl' if dist.is_nccl_available else 'gloo')->
train->
callbacks.run('on_pretrain_routine_start')->
model=Model(cfg or ckpt['model'].yaml,ch=3,nc=nc,anchors=hyp.get('anchor')),
to(device)->
= DetectionModel->model=parse_model(yaml)->
= initialize_weights()->
freeze->
optimizer=smart_optimizer(model,opt.optimizer,hyp['lr0'],hyp['momentum'],
hyp['weight_decay'])->
scheduler=lr_scheduler.LambdaLR(optimizer)->
ema=ModelEMA(model)->
train_loader,dataset=create_dataloader(train_pathm...)->
callbacks.run('on_pretrain_routine_end')->
stopper,stop=EarlyStopping()->
compute_loss=ComputeLoss(model)->
callbacks.run('on_train_start')->
callbacks.run('on_train_epoch_start')->
model.train()->
optimizer.zero_grad()->
callbacks.run('on_train_batch_start')->
pred=model(imgs)->
loss,loss_items=compute_loss(pred,target)->
= tcls,tbox,indices,anchor=build_targets(p,target)->
loss.backward()->
callbacks.run('on_train_batch_end')->
scheduler.step()->
callbacks.run('on_train_epoch_end')

4.1 在data下新建一个person.yaml

# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..]
path: /home/imcs/local_disk/D0011  # dataset root dir
train: train_yolo.txt  # train images (relative to 'path') 118287 images
val: val_yolo.txt  # val images (relative to 'path') 5000 images
test: test_yolo.txt  # 20288 of 40670 images, submit to https://competitions.codalab.org/competitions/20794

# Classes
nc: 1  # number of classes
names: ['person']  # class names

4.2 在models更新yolov5s.yaml中，把nc改成类别。

4.3 分布式训练

改一下代码中的device参数，0,1,2,3

python -m torch.distributed.launch --nproc_per_node 4 train.py

batch_size=32，yolov5的代码这块写的很好，即便在k80上也不需要很大的改动。

在8卡k80上，上面这种写法出现了问题，在4卡k80上没有错，很奇怪。

在train.py中添加：

import torch.distributed as dist 
dist.init_process_group(backend='gloo', init_method='env://')

 在527行注释掉，不让dist初始化两次

# dist.init_process_group(backend="nccl" if dist.is_nccl_available() else "gloo")

在utils的torch_utils.py中

@contextmanager
def torch_distributed_zero_first(local_rank: int):
    """
    Decorator to make all processes in distributed training wait for each local_master to do something.
    """
    if local_rank not in [-1, 0]:
        torch.distributed.barrier()
    yield
    if local_rank == 0:
        torch.distributed.barrier()

换成torch.distibuted.barrier()

最终运行：

python -m torch.distributed.launch --nproc_per_node=8 --nnodes=1 --node_rank=0 --master_addr=localhost --master_port=22222 train.py 

5. 测试

detect.py

weights/source(支持视频图片url图片集合)/data/imgsz...->
device=select_device()->
model=DetectMultiBackend(weights)->
imgsz=check_img_size(imgsz,stride)->
dataset=LoadImages()->
model.warmup()->
for path,im,im0s,vid_cap,s in dataset->
= path=files[count]->
= cv2.imread/cap_read()->
= letterbox(img0,img_size,stride,auto)[0]->
= img=img.transpose((2,0,1))[::-1]->
= img=np.ascontiguousarray(img)->
pred=model(im,augment)->
pred=non_max_suppression(pred,conf_thres,iou_thres,classes,agnostic_nms,max_det)->
annotator=Annotator()->
annotator.box_label()->
annotator.result()

300轮：metrics/precision：0.83757，metrics/recall:0.749，map_0.5:0.84328，map_0.5:0.95:0.56934

6.部署(c++)

windows/linux(x86)/linux(arm)树莓派

libtorch/openvino/ncnn/onnxruntime/rknn