标牌工作手册_标牌手册-优快云博客

本文介绍了一种交通标志数据集的标注方法及基于YOLOv4的模型训练流程，包括数据预处理步骤、模型配置文件调整、训练与测试指令等关键信息。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

一、数据集

（一）分类

00最大限速牌，000005到000140：最大限速5到最大限速140
01结束最大限速牌，010005到010140：结束最大限速5到结束最大限速140
02最大轴重牌，020020：最大轴重2t
03限重，030010：限重1t
04限高，040010：限高1m
05限宽，050010：限宽1m
06最小限速，060005：最小限速5
07禁止标牌，070004：禁止三轮车
08信息指示牌，080004：机动车左转
09警告牌，094003：注意儿童
0A彩色标牌，0A0001：龙门架方形大牌子蓝色
0B可变标牌，0B0005：可变标牌5
0C诱导标志，0C0001:诱导标志

标注的文档更新在这里：
https://10.60.158.228:8443/svn/PCM-ProjectConfigurationManagement/04 Department Library(部门库)/03 Map Product Center(地图产品中心)/07 Ark Project Department(方舟项目部)/05 Projects(项目资料)/2021年/样本标注/Sprint3
在这里插入图片描述

（二）工作空间的notebook

HAD-slp6562
navinfo

（三）数据预处理

1、用VSCode自带的python编译器，执行step2_json2xml.py，会报错误UnicodeEncodeError

解决办法：用notebook里的python编译器

2、直接从AI平台挂载的数据集，普遍存在图像后缀名为JPG的

解决办法：取消注释step2_json2xml.py中重命名部分代码

3、构建数据集时没有选择过滤，有可能出现KeyError: 'w’的报错

原因是，报错的json文件中，存在没有“w” key的分割要素。
解决办法：1、构建数据集时，选择过滤
2、修改/opt/conda/lib/python3.6/site-packages/aicloud/json_to_voc.py代码，修改部分的代码如下：
在这里插入图片描述

4、step4_xml2txt.py和交通灯不同，修改如下：

import xml.etree.ElementTree as ET
import pickle
import os
from os import listdir, getcwd
from os.path import join

sets=[('2007', 'train'), ('2007', 'val'), ('2007', 'test')]

classes = ["00","01","02","03","04","05","06","07","08","09","0A","0B","0C"]

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(year, image_id):
    if os.path.exists('/root/data_sign/sign_test/VOCdevkit/VOC%s/Annotations/%s.xml'%(year, image_id)):
        in_file = open('/root/data_sign/sign_test/VOCdevkit/VOC%s/Annotations/%s.xml'%(year, image_id))
        out_file = open('/root/data_sign/sign_test/VOCdevkit/VOC%s/labels/%s.txt'%(year, 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
            cls = obj.find('name').text
            if cls in classes:
                cls_id = classes.index(cls)
            else:
                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(cls_id) + " " + " ".join([str(a) for a in bb]) + '\n')

wd = getcwd()

for year, image_set in sets:
    if not os.path.exists('/root/data_sign/sign_test/VOCdevkit/VOC%s/labels/'%(year)):
        os.makedirs('/root/data_sign/sign_test/VOCdevkit/VOC%s/labels/'%(year))
    image_ids = open('/root/data_sign/sign_test/VOCdevkit/VOC%s/ImageSets/Main/%s.txt'%(year, image_set)).read().strip().split()
    list_file = open('/root/data_sign/sign_test/VOCdevkit/%s_%s.txt'%(year, image_set), 'w')
    i = 0
    for image_id in image_ids:
        i = i+1
        # convert_annotation(year, image_id)
        # print(i,"sucessfully")
        list_file.write('/root/data_sign/sign_test/VOCdevkit/VOC2007/JPEGImages/%s.jpg\n'%(image_id))
        
    list_file.close()

os.system("cat /root/data_sign/sign_test/VOCdevkit/2007_train.txt /root/data_sign/sign_test/VOCdevkit/2007_val.txt /root/data_sign/sign_test/VOCdevkit/2007_test.txt > /root/data_sign/sign_test/VOCdevkit/train_all.txt")

（四）模型训练

1、crop.names的修改

2、crop.data的修改

3、crop.cfg的修改

3.1 修改anchor

./darknet detector calc_anchors crop/crop.data -num_of_clusters 9 -width 800 -height 800

3.2 修改类别

同时修改filters和class,filters = 3 * (class+5)，有三处

4、开始训练

./darknet detector train crop/crop.data crop/crop.cfg backup/yolov4.conv.137 -gpus 0,1 -dont_show -mjpeg_port 8092 -map

./darknet detector train crop/crop.data crop/crop.cfg backup/yolov4.conv.137 -gpus 0 -dont_show -mjpeg_port 8092 -map

./darknet detector train /root/work/darknet/cfg/NI.data /root/work/darknet/cfg/yolov4-ni.cfg backup/yolov4.conv.137 -gpus 0 -dont_show -mjpeg_port 8092 -map

./darknet detector train crop/crop.data crop/crop.cfg backup/yolov4-ni_best.weights -gpu 0 -dont_show -mjpeg_port 8092 -map -clear

5、测试

按照比例缩放（留白缩放）的推理代码：

./darknet detector test /root/work/darknet/cfg/NI.data /root/work/darknet/cfg/yolov4-ni.cfg /root/work/darknet/backup/yolov4-ni_best.weights  -letter_box -ext_output  -dont_show -out /root/work/result.json < /root/work/cp_test.txt

./darknet detector map /root/work/darknet/cfg/NI.data /root/work/darknet/cfg/yolov4-ni.cfg /root/work/darknet/backup/yolov4-ni_best.weights  -iou_thresh 0.5