基于PaddleX实现电梯电瓶车检测

源地址.

一、项目背景

   电动车因为具有便捷、轻巧、无污染等优点，成为了广大民众出行的重要选择，但是电动车自身的电瓶存在安全隐患，目前仍然没有办法实现绝对的安全，近年来，电动车在室内、电梯内发生自燃、自爆的事件在新闻上常有报道，这些现象严重危害到了居民的生命、财产安全。
   因此，本项目将以“电梯电瓶车检测”为首要目标进行开发，待项目开发成功以后，可以很容易地将本项目迁移到居民楼门口、办公楼门口、餐厅门口等众多关键场地，而且本项目的成本低廉，检测效率高，可应用的场景十分广泛，具有巨大的商业价值。

二、数据集简介

1. 数据获取

• 为了获取大量数据，以及节省标注时间，所以将pascal-voc和coco2017中的摩托车数据提取出来，作为数据集。

LabelMe使用方法

2. 解压数据集和安装第三方库

!rm -rf data/pascalvoc
!rm -rf data/val2017
!rm -rf data/train2017
!rm -rf data/annotations
!unzip -oq /home/aistudio/data/data4379/pascalvoc.zip -d data/
!unzip -oq /home/aistudio/data/data7122/train2017.zip -d data/
!unzip -oq /home/aistudio/data/data7122/val2017.zip -d data/
!unzip -oq /home/aistudio/data/data7122/annotations_trainval2017.zip -d data/
!rm -rf motorbike

!pip -q install lxml

3. COCO2VOC

• 转换代码来源

import os
import json
import cv2
from lxml import etree
import xml.etree.cElementTree as ET
import time
import pandas as pd
from tqdm import tqdm
import json


def coco2voc(anno, xml_dir):
    if not os.path.exists(xml_dir):
        os.makedirs(xml_dir)

    with open(anno, 'r', encoding='utf-8') as load_f:
        f = json.load(load_f)

    imgs = f['images']

    df_cate = pd.DataFrame(f['categories'])
    _ = df_cate.sort_values(["id"],ascending=True)
    df_anno = pd.DataFrame(f['annotations'])
    categories = dict(zip(df_cate.id.values, df_cate.name.values))

    for i in tqdm(range(len(imgs))):
        xml_content = []
        file_name = imgs[i]['file_name']
        height = imgs[i]['height']
        img_id = imgs[i]['id']
        width = imgs[i]['width']

        xml_content.append("<annotation>")
        xml_content.append("	<folder>VOC2007</folder>")
        xml_content.append("	<filename>"+file_name+"</filename>")
        xml_content.append("	<size>")
        xml_content.append("		<width>"+str(width)+"</width>")
        xml_content.append("		<height>"+str(height)+"</height>")
        xml_content.append("		<depth>3</depth>")
        xml_content.append("	</size>")
        xml_content.append("	<segmented>0</segmented>")
        
        annos = df_anno[df_anno["image_id"].isin([img_id])]

        for index, row in annos.iterrows():
            bbox = row["bbox"]
            category_id = row["category_id"]
            cate_name = categories[category_id]

            
            xml_content.append("	<object>")
            xml_content.append("		<name>"+cate_name+"</name>")
            xml_content.append("		<pose>Unspecified</pose>")
            xml_content.append("		<truncated>0</truncated>")
            xml_content.append("		<difficult>0</difficult>")
            xml_content.append("		<bndbox>")
            xml_content.append("			<xmin>"+str(int(bbox[0]))+"</xmin>")
            xml_content.append("			<ymin>"+str(int(bbox[1]))+"</ymin>")
            xml_content.append("			<xmax>"+str(int(bbox[0]+bbox[2]))+"</xmax>")
            xml_content.append("			<ymax>"+str(int(bbox[1]+bbox[3]))+"</ymax>")
            xml_content.append("		</bndbox>")
            xml_content.append("	</object>")
        xml_content.append("</annotation>")

        x = xml_content
        xml_content=[x[i] for i in range(0,len(x)) if x[i]!="\n"]


        xml_path = os.path.join(xml_dir,file_name.split('.')[-2] + '.xml')

        with open(xml_path, 'w+',encoding="utf8") as f:
            f.write('\n'.join(xml_content))
        xml_content[:]=[]

• 转换

coco2voc('data/annotations/instances_train2017.json', 'data/train2017voc/Annotations')
coco2voc('data/annotations/instances_val2017.json', 'data/val2017voc/Annotations')

4. 将pascal-voc和coco2017中的摩托车数据提取出来

• 需要将coco2017中的motorcycle转换成motorbike

import os
import shutil
from tqdm import tqdm
import cv2


def voc2sc(ann_filepath, img_filepath, img_savepath, ann_savepath, classes, sc_class):
    """
    ann_filepath: xml文件夹路径
    img_filepath: 原图文件夹路径
    img_savepath: 保存原图路径
    ann_savepath: 保存xml文件夹路径
    classes: 所有类别
    sc_class: 挑选类别
    """
    if not os.path.exists(img_savepath):
        os.makedirs(img_savepath)
    if not os.path.exists(ann_savepath):
        os.makedirs(ann_savepath)

    names = locals()

    for file in tqdm(os.listdir(ann_filepath)):
        fp = open(ann_filepath + '/' + file)
        ann_savefile=ann_savepath+file
        fp_w = open(ann_savefile, 'w')
        lines = fp.readlines()
        ind_start = []
        ind_end = []
        lines_id_start = lines[:]
        lines_id_end = lines[:]
        #在xml中找到object块，并将其记录下来
        while "\t<object>\n" in lines_id_start:
            a = lines_id_start.index("\t<object>\n")
            ind_start.append(a)
            lines_id_start[a] = "delete"
        while "\t</object>\n" in lines_id_end:
            b = lines_id_end.index("\t</object>\n")
            ind_end.append(b)
            lines_id_end[b] = "delete"
        #names中存放所有的object块
        i = 0
        for k in range(0, len(ind_start)):
            names['block%d' % k] = []
            for j in range(0, len(classes)):
                if classes[j] in lines[ind_start[i] + 1]:
                    a = ind_start[i]
                    for o in range(ind_end[i] - ind_start[i] + 1):
                        names['block%d' % k].append(lines[a + o])
                    break
            i += 1
            #print(names['block%d' % k])
        # xml头
        if len(ind_start) == 0:
            fp.close()
            fp_w.close()
            os.remove(ann_savefile)
            continue
        string_start = lines[0:ind_start[0]]
        #xml尾
        string_end = [lines[len(lines) - 1]]
        #在给定的类中搜索，若存在则，写入object块信息
        a = 0
        for k in range(0, len(ind_start)):
            if sc_class in names['block%d' % k]:
                a += 1
                string_start += names['block%d' % k]
        string_start += string_end
        if 'size' not in str(string_start) and a != 0:
            s = """    <size>
        <width>{}</width>
        <height>{}</height>
        <depth>3</depth>
    </size>
"""
            img = cv2.imread(img_filepath + os.path.splitext(file)[0] + ".jpg")
            idx = string_start.index('\t<object>\n')
            string_start.insert(idx, s.format(img.shape[1], img.shape[0]))
        for c in range(0, len(string_start)):
            if '\t\t<name>motorcycle</name>\n' == string_start[c]:
                string_start[c] = '\t\t<name>motorbike</name>\n'
            fp_w.write(string_start[c])
        fp_w.close()

        #如果没有我们寻找的模块，则删除此xml，有的话拷贝图片
        if a == 0:
            os.remove(ann_savefile)
        else:
            name_img = img_filepath + os.path.splitext(file)[0] + ".jpg"
            shutil.copy(name_img, img_savepath)
        fp.close()

• 提取pascalvoc中的摩托车

# pascalvoc类别
pascalvoc_classes = ['aeroplane','bicycle','bird', 'boat', 'bottle',
                    'bus', 'car', 'cat', 'chair', 'cow','diningtable',
                    'dog', 'horse', 'motorbike', 'pottedplant',
                    'sheep', 'sofa', 'train', 'tvmonitor', 'person']

# pascalvoc挑选类别
pascalvoc_class = '\t\t<name>motorbike</name>\n'

# 转换
voc2sc('data/pascalvoc/VOCdevkit/VOC2007/Annotations/', 'data/pascalvoc/VOCdevkit/VOC2007/JPEGImages/',
        'motorbike/JPEGImages/', 'motorbike/Annotations/',
        pascalvoc_classes, pascalvoc_class)
voc2sc('data/pascalvoc/VOCdevkit/VOC2012/Annotations/', 'data/pascalvoc/VOCdevkit/VOC2012/JPEGImages/',
        'motorbike/JPEGImages/', 'motorbike/Annotations/',
        pascalvoc_classes, pascalvoc_class)

with open('motorbike/labels.txt', 'w') as f: 
    f.write('motorbike\n')

print('len datas:', len(os.listdir('motorbike/JPEGImages')))

• 提取coco2017中的摩托车

# coco2017类别
coco2017_classes = []
with open('coco_category.txt') as f:
    for line in f.readlines():
        coco2017_classes.append(line.strip())

# coco2017挑选类别
coco2017_class = '\t\t<name>motorcycle</name>\n'

# 转换
voc2sc('data/train2017voc/Annotations/', 'data/train2017/',
        'motorbike/JPEGImages/', 'motorbike/Annotations/',
        coco2017_classes, coco2017_class)
voc2sc('data/val2017voc/Annotations/', 'data/val2017/',
        'motorbike/JPEGImages/', 'motorbike/Annotations/',
        coco2017_classes, coco2017_class)

with open('motorbike/labels.txt', 'w') as f: 
    f.write('motorbike\n')

print('len datas:', len(os.listdir('motorbike/JPEGImages')))

5. 将文件路径写入.txt

import os
from tqdm import tqdm
import numpy

def path2txt(images_dir, anno_dir, txt_save_path, split_to_eval=False):
    data = []
    for line in os.listdir(images_dir):
        data.append([os.path.join(images_dir, line), os.path.join(anno_dir, line.replace('jpg', 'xml'))])
    
    if split_to_eval:
        numpy.random.shuffle(data)
        eval_data = data[:int(len(data)/10*2)]
        data = data[int(len(data)/10*2):]
        with open(txt_save_path.replace('train', 'val'), 'w') as f:
            for item in tqdm(eval_data):
                f.write('{} {}\n'.format(item[0], item[1]))

    with open(txt_save_path, 'w') as f:
        for item in tqdm(data):
            f.write('{} {}\n'.format(item[0], item[1]))


path2txt('motorbike/JPEGImages', 'motorbike/Annotations', 'motorbike/train.txt', split_to_eval=True)

三、模型选择与开发

1. 安装paddlex

paddlex使用文档

!pip -q install paddlex==1.3

2. 数据加载与预处理

import os
os.environ['CUDA_VISIBLE_DEVICES'] = '0'

from paddlex.det import transforms
import paddlex as pdx


# 定义训练和验证时的transforms
train_transforms = transforms.Compose([
    transforms.MixupImage(mixup_epoch=250),
    transforms.RandomDistort(),
    transforms.RandomExpand(),
    transforms.RandomCrop(),
    transforms.Resize(target_size=608, interp='RANDOM'),
    transforms.RandomHorizontalFlip(),
    transforms.Normalize(),
])

eval_transforms = transforms.Compose([
    transforms.Resize(target_size=608, interp='CUBIC'),
    transforms.Normalize(),
])

# 定义训练和验证所用的数据集
train_dataset = pdx.datasets.VOCDetection(
    data_dir='./',
    file_list='motorbike/train.txt',
    label_list='motorbike/labels.txt',
    transforms=train_transforms,
    shuffle=True)

eval_dataset = pdx.datasets.VOCDetection(
    data_dir='./',
    file_list='motorbike/val.txt',
    label_list='motorbike/labels.txt',
    transforms=eval_transforms)

3. 模型选择

• 这里我们选择的模型是YOLOv3-MobileNetV1，其适合于后期边缘端的部署。
• 详细信息如下：

模型	模型大小	预测时间(ms/image）	BoxAP（%）
YOLOv3-MobileNetV1	99.2MB	11.834	9.3

4. 模型训练

• 注：有时候会报这样的错，请等notebook空闲几十分钟（刚转换完数据会这样）。
  

# 初始化模型，并进行训练
num_classes = len(train_dataset.labels)
batch_size = 64

model = pdx.det.YOLOv3(num_classes=num_classes, backbone='MobileNetV1')


model.train(
    num_epochs=200,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    train_batch_size=batch_size,
    learning_rate=1e-4,
    warmup_steps=len(train_dataset.file_list)//2,
    save_interval_epochs=50,
    save_dir='output/yolov3_mobilenetv1',
    metric='VOC',
    use_vdl=True,
    early_stop=True)

四、效果展示

1. 模型预测

from paddlex.det import transforms
import paddlex as pdx
import os


# 数据处理
eval_transforms = transforms.Compose([
    transforms.Resize(target_size=608, interp='CUBIC'),
    transforms.Normalize(),
])
# 模型载入
model = pdx.load_model('output/yolov3_mobilenetv1/best_model')
# 结果可视化
for p in os.listdir('example/images'):
    img_path = 'example/images/' + p
    result = model.predict(img_file=img_path, transforms=eval_transforms)
    pdx.det.visualize(image=img_path, result=result, threshold=0.1, save_dir='./visualize')

2. 预测结果可视化

3. 部署模型导出

部署模型导出

!paddlex --export_inference --model_dir=output/yolov3_mobilenetv1/best_model --save_dir=inference_model --fixed_input_shape=[608,608]

五、总结与升华

• 1. 由于未拿到硬件，部署部分等待更新~
• 2. 目前的电瓶车数据较少，需要用LabelMe标注更多的电瓶车数据。
• 3. 更多详细paddlex请访问PaddleX。

个人简介

• 菜鸡一枚~，啥都不会，干饭第一！！！
• 我在AI Studio上获得钻石等级，点亮8个徽章，来互关呀~ https://aistudio.baidu.com/aistudio/personalcenter/thirdview/517701