基於樹莓派與YOLOv3模型的人體目標檢測小車（二）

上篇文章介紹了如何搭建深度學習環境，在Ubuntu18.04TLS上搭建起了 CUDA：9.0+cuDNN7.0+tensorflow-gpu 1.9 的訓練環境。本篇文章將介紹如何制作自己的數據集，並訓練模型。

本文訓練數據集包括從VOC數據集中提取出6095張人體圖片，以及使用LabelImg工具標注的200張python爬蟲程序獲取的人體圖片作為補充。

一、爬取人體圖片並標記

# coding=utf-8
"""根據搜索詞下載百度圖片"""
import re
import sys
import urllib
import requests


def getPage(keyword, page, n):
    page = page * n
    keyword = urllib.parse.quote(keyword, safe='/')
    url_begin = "http://image.baidu.com/search/flip?tn=baiduimage&ie=utf-8&word="
    url = url_begin + keyword + "&pn=" + str(page) + "&gsm=" + str(hex(page)) + "&ct=&ic=0&lm=-1&width=0&height=0"
    return url


def get_onepage_urls(onepageurl):
    try:
        html = requests.get(onepageurl).text
    except Exception as e:
        print(e)
        pic_urls = []
        return pic_urls
    pic_urls = re.findall('"objURL":"(.*?)",', html, re.S)
    return pic_urls


def down_pic(pic_urls):
    """給出圖片鏈接列表, 下載所有圖片"""
    for i, pic_url in enumerate(pic_urls):
        try:
            pic = requests.get(pic_url, timeout=15)
            string = str(i + 1) + '.jpg'
            with open(string, 'wb') as f:
                f.write(pic.content)
                print('成功下載第%s張圖片: %s' % (str(i + 1), str(pic_url)))
        except Exception as e:
            print('下載第%s張圖片時失敗: %s' % (str(i + 1), str(pic_url)))
            print(e)
            continue


if __name__ == '__main__':
    keyword = '行人圖片'  # 關鍵詞, 改為你想輸入的詞即可, 相當於在百度圖片里搜索一樣
    page_begin = 0
    page_number = 100
    image_number = 3
    all_pic_urls = []
    while 1:
        if page_begin > image_number:
            break
        print("第%d次請求數據", [page_begin])
        url = getPage(keyword, page_begin, page_number)
        onepage_urls = get_onepage_urls(url)
        page_begin += 1

        all_pic_urls.extend(onepage_urls)

    down_pic(list(set(all_pic_urls)))

使用labelimg標記圖片

二、從VOC數據集里提取出人體圖片

import os
import os.path
import shutil

fileDir_ann = "D:\\VOC\\VOCdevkit\\VOC2012\\Annotations"
fileDir_img = "D:\\VOC\\VOCdevkit\\VOC2012\\JPEGImages\\"
saveDir_img = "D:\\VOC\\VOCdevkit\\VOC2012\\JPEGImages_ssd\\"

if not os.path.exists(saveDir_img):
    os.mkdir(saveDir_img)

names = locals()

for files in os.walk(fileDir_ann):
    for file in files[2]:



        saveDir_ann = "D:\\VOC\\VOCdevkit\\VOC2012\\Annotations_ssd\\"
        if not os.path.exists(saveDir_ann):
            os.mkdir(saveDir_ann)

        fp = open(fileDir_ann + '\\' + file)
        saveDir_ann = saveDir_ann + file
        fp_w = open(saveDir_ann, 'w')
        classes = ['aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', '>cat<', 'chair', 'cow',
                   'diningtable', \
                   'dog', 'horse', 'motorbike', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor', 'person']

        lines = fp.readlines()

        ind_start = []
        ind_end = []
        lines_id_start = lines[:]
        lines_id_end = lines[:]

        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"

        i = 0
        for k in range(0, len(ind_start)):
            for j in range(0, len(classes)):
                if classes[j] in lines[ind_start[i] + 1]:
                    a = ind_start[i]
                    names['block%d' % k] = [lines[a], lines[a + 1], \
                                            lines[a + 2], lines[a + 3], lines[a + 4], \
                                            lines[a + 5], lines[a + 6], lines[a + 7], \
                                            lines[a + 8], lines[a + 9], lines[a + 10], \
                                            lines[ind_end[i]]]
                    break
            i += 1

        classes1 = '\t\t<name>person</name>\n'



        string_start = lines[0:ind_start[0]]
        string_end = [lines[len(lines) - 1]]

        a = 0
        for k in range(0, len(ind_start)):
            if classes1 in names['block%d' % k]:
                a += 1
                string_start += names['block%d' % k]



        string_start += string_end
        for c in range(0, len(string_start)):
            fp_w.write(string_start[c])
        fp_w.close()

        if a == 0:
            os.remove(saveDir_ann)
        else:
            name_img = fileDir_img + os.path.splitext(file)[0] + ".jpg"
            shutil.copy(name_img, saveDir_img)
        fp.close()

三、修改YOLOv3 tiny 配置文件

• yolov3-tiny.cfg

batch = 64

max_batchs=500200 迭代次數

learning_rate = 0.001

steps = 400000,450000 scales =.1,.1 學習率在400000和450000次時縮小10倍

class = 1 設置單類別

• 刪除voc.names中其余名字，只保留person
• 修改voc.data中classes值為1

四、下載預訓練權重開始訓練

預訓練權重可以減少前期的迭代次數，加速訓練過程。

wget https://pjreddie.com/media/files/darknet53.conv.74

開始訓練：

./darknet detector train cfg/voc.data cfg/yolov3-voc-tiny.cfg darknet53.conv.74

通過繪制訓練過程的loss曲線可知，開始時loss下降較快，之后開始在一水平線上波動。

訓練結束得到yolov3-voc_final.weights模型文件。