基于EdgeBoard飞浆开发套件和PicoDet部署夜间场景检测模型

在夜间微弱的灯光下，摄像头拍摄的画质非常差，基于夜间场景的识别模型训练有助于夜间安防。本项目基于ExDark数据集、PicoDet、EdgeBoard飞桨开发套件和PyQt5进行夜间场景特定物品检测模型的训练和部署。

材料准备

1. ExDark数据集
2. Aistudio在线运行环境
3. EdgeBoard飞桨开发套件（一个板子，可以用本地CPU代替~）
4. 摄像头（USB摄像头，尽可能便宜即可）

代码

训练PicoDet模型

基础环境准备

! unzip -oq data/data129450/ExDark.zip

# 克隆PaddleDetection仓库
import os
if not os.path.exists('PaddleDetection'):
    !git clone https://github.com/PaddlePaddle/PaddleDetection.git

# 安装其他依赖
%cd PaddleDetection
! pip install -r requirements.txt

# 编译安装paddledet
! python setup.py install

# 测试一下PaddleDet环境有没有准备好
! python ppdet/modeling/tests/test_architectures.py

调整数据格式

将数据格式调整为voc格式，并且准备paddledet需要的文件

正所谓是闻名不如见面吗，不熟悉voc格式的小伙伴可以通过以下语句查看一个voc格式的示例~

%cd ~
! wget https://paddlemodels.bj.bcebos.com/object_detection/roadsign_voc.tar
! tar -xvf /home/aistudio/roadsign_voc.tar

# ExDark转voc
%cd ~

# thanks to https://blog.csdn.net/hu694028833/article/details/81089959
import xml.dom.minidom as minidom
import cv2

def convert2xml(img_dir,annotation_dir,save_dir):
    img = cv2.imread(img_dir)
    h,w,c = img.shape

    dom = minidom.getDOMImplementation().createDocument(None,'annotations',None)
    root = dom.documentElement

    element = dom.createElement('filename')
    element.appendChild(dom.createTextNode(img_dir.split('/')[-1]))
    root.appendChild(element)

    element = dom.createElement('size')
    element_c = dom.createElement('width')
    element_c.appendChild(dom.createTextNode(str(w)))
    element.appendChild(element_c)
    element_c = dom.createElement('height')
    element_c.appendChild(dom.createTextNode(str(h)))
    element.appendChild(element_c)
    element_c = dom.createElement('depth')
    element_c.appendChild(dom.createTextNode(str(c)))
    element.appendChild(element_c)
    root.appendChild(element)

    with open(annotation_dir) as f:
        for line in f.readlines():
            if '% bbGt' in line:
                continue
            else:
                element = dom.createElement('object')

                name = line.split(' ')[0]
                xmin = int(line.split(' ')[1])
                ymin = int(line.split(' ')[2])
                xlen = int(line.split(' ')[3])
                ylen = int(line.split(' ')[4])

                element_c = dom.createElement('name')
                element_c.appendChild(dom.createTextNode(name))
                element.appendChild(element_c)
                element_c = dom.createElement('bndbox')
                
                element.appendChild(element_c)
                element_cc = dom.createElement('xmin')
                element_cc.appendChild(dom.createTextNode(str(xmin)))
                element_c.appendChild(element_cc)
                element_cc = dom.createElement('ymin')
                element_cc.appendChild(dom.createTextNode(str(ymin)))
                element_c.appendChild(element_cc)
                element_cc = dom.createElement('xmax')
                element_cc.appendChild(dom.createTextNode(str(xmin+xlen)))
                element_c.appendChild(element_cc)
                element_cc = dom.createElement('ymax')
                element_cc.appendChild(dom.createTextNode(str(ymin+ylen)))
                element_c.appendChild(element_cc)

                root.appendChild(element)

    with open(save_dir, 'w', encoding='utf-8') as f:
        dom.writexml(f, addindent='\t', newl='\n',encoding='utf-8')

convert2xml('ExDark/images/Bicycle/2015_00001.png','ExDark/Annnotations/Bicycle/2015_00001.png.txt','test.xml')

/home/aistudio

# thanks to https://www.runoob.com/python/os-walk.html
# 用一个train.txt记录数据
%cd ~
import os
! mkdir -p ExDark/annotations
with open('train.txt','w') as f:
    for root, dirs, files in os.walk("ExDark/images", topdown=False):
        for name in files:
            img_dir = os.path.join(root, name)
            if 'ipynb_checkpoints' in img_dir:
                continue
            annotation_dir = img_dir.replace('images','Annnotations')+'.txt'
            save_dir = 'ExDark/annotations/'+img_dir.split('/')[-1]+'.xml'
            try:
                convert2xml(img_dir,annotation_dir,save_dir)
            except:
                continue
            f.write(img_dir+' '+save_dir+'\n')

# 准备标签文件
%cd ~
with open('label_list.txt','w') as f:
    for item in os.listdir("ExDark/images"):
        f.write(item+'\n')

/home/aistudio

准备yml文件

在PaddleDetection/configs/picodet/目录下准备一个名为myPicodetXsVoc.yml的文件，内容为

_BASE_: [
  '../datasets/voc.yml',
  '../runtime.yml',
  '_base_/picodet_v2.yml',
  '_base_/optimizer_300e.yml',
  '_base_/picodet_320_reader.yml',
]

pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/LCNet_x0_35_pretrained.pdparams
weights: output/picodet_xs_320_coco/best_model
find_unused_parameters: True
use_ema: true
epoch: 8
snapshot_epoch: 2

LCNet:
  scale: 0.35
  feature_maps: [3, 4, 5]

LCPAN:
  out_channels: 96

PicoHeadV2:
  conv_feat:
    name: PicoFeat
    feat_in: 96
    feat_out: 96
    num_convs: 2
    num_fpn_stride: 4
    norm_type: bn
    share_cls_reg: True
    use_se: True
  feat_in_chan: 96

TrainReader:
  batch_size: 64

LearningRate:
  base_lr: 0.32
  schedulers:
  - !CosineDecay
    max_epochs: 300
  - !LinearWarmup
    start_factor: 0.1
    steps: 300

将PaddleDetection/configs/datasets/voc.yml修改为如下内容

metric: VOC
map_type: 11point
num_classes: 12

TrainDataset:
  !VOCDataSet
    dataset_dir: /home/aistudio
    anno_path: /home/aistudio/train.txt
    label_list: /home/aistudio/label_list.txt
    data_fields: ['image', 'gt_bbox', 'gt_class', 'difficult']

EvalDataset:
  !VOCDataSet
    dataset_dir: /home/aistudio
    anno_path: /home/aistudio/train.txt
    label_list: /home/aistudio/label_list.txt
    data_fields: ['image', 'gt_bbox', 'gt_class', 'difficult']

TestDataset:
  !ImageFolder
    anno_path: /home/aistudio/label_list.txt

训练

10 epoch 24 min

%cd ~/PaddleDetection
! export CUDA_VISIBLE_DEVICES=0 #windows和Mac下不需要执行该命令
! python tools/train.py -c configs/picodet/myPicodetXsVoc.yml

预测

预测一下看看结果吧，如果啥也没检测到就修改threshold，毕竟是黑暗场景~调低一点就能看到了。

结果展示：

# 预测代码
%cd ~/PaddleDetection
! export CUDA_VISIBLE_DEVICES=0 #windows和Mac下不需要执行该命令
! python tools/infer.py -c configs/picodet/myPicodetXsVoc.yml \
                    --infer_img=/home/aistudio/ExDark/images/Bicycle/2015_00001.png \
                    --output_dir=infer_output/ \
                    --draw_threshold=0.5 \
                    -o weights=output/myPicodetXsVoc/model_final \
                    --use_vdl=False

/home/aistudio/PaddleDetection
W0910 18:38:48.325160 14436 gpu_resources.cc:61] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.2, Runtime API Version: 11.2
W0910 18:38:48.328585 14436 gpu_resources.cc:91] device: 0, cuDNN Version: 8.2.
[09/10 18:38:48] ppdet.utils.checkpoint INFO: Finish loading model weights: output/myPicodetXsVoc/model_final.pdparams
100%|█████████████████████████████████████████████| 1/1 [00:01<00:00,  1.99s/it]
[09/10 18:38:50] ppdet.engine INFO: Detection bbox results save in infer_output/2015_00001.png

模型导出

%cd ~/PaddleDetection
! python tools/export_model.py -c configs/picodet/myPicodetXsVoc.yml \
                               --output_dir=./inference_model \
                               -o weights=output/myPicodetXsVoc/model_final \
                               TestReader.inputs_def.image_shape=[3,320,320]

/home/aistudio/PaddleDetection
[09/06 20:38:43] ppdet.utils.checkpoint INFO: Finish loading model weights: output/myPicodetXsVoc/model_final.pdparams
[09/06 20:38:43] ppdet.engine INFO: Export inference config file to ./inference_model/myPicodetXsVoc/infer_cfg.yml
[09/06 20:38:48] ppdet.engine INFO: Export model and saved in ./inference_model/myPicodetXsVoc

部署

模型转化

Fork项目EdgeBoard飞桨开发套件，根据项目内说明，执行即可。

需要注意以下两点：

• EdgeBoard飞桨开发套件中除章节“模型推理”外，均可以在Aistudio上执行
• 如果Fork上述项目，在执行过程中不顺利，可以尝试在Terminal中运行。

部署

转化后的模型为一个tar文件。

将转化后的.tar文件放在EdgeBoard指定路径下，并修改对应Config信息即可（具体参考EdgeBoard飞桨开发套件）。

具体的EB推理代码如下（由于使用了pyqt5和摄像头需要在本地运行）：

#!/usr/bin/python3
# -*- coding: utf-8 -*-

"""
Thanks to 
zetcode.com
https://blog.csdn.net/u014453898/article/details/88083173
https://github.com/maicss/PyQt-Chinese-tutorial
https://maicss.gitbooks.io/pyqt5/content/
https://blog.csdn.net/m0_37811342/article/details/108741505
"""

import sys
import cv2
import numpy as np
import math
import time
import collections
import os
import sys
from PyQt5.QtWidgets import QWidget, QDesktopWidget, QApplication, QPushButton, QFileDialog, QLabel, QTextEdit, \
    QGridLayout, QFrame, QColorDialog
from PyQt5.QtCore import QTimer, QUrl
from PyQt5.QtGui import QColor, QImage, QPixmap

"""Demo for ppnc runtime on board"""
import json

# add python path of ppnc/python to sys.path
parent_path = os.path.abspath(os.path.join(__file__, *(['..'] * 2)))
sys.path.insert(0, parent_path + '/python')

from ppnc import PPNCPredictor

def parse_config(config_file):
    """parse config"""
    with open(config_file, "r") as f:
        config = json.load(f)
    return config

def preprocess(img, input_size=(320, 320)):
    """parse image

    Args:
        img_path (str): path
        input_size (tuple, optional): h x w. Defaults to (320, 320).
    """
    input = {}
    # img = cv2.imread(img_path)
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    
    scale_factor = [input_size[0] / img.shape[0], input_size[1] / img.shape[1]]
    factor = np.array(scale_factor, dtype=np.float32)
    input['scale_factor'] = factor.reshape((1, 2))
    img = cv2.resize(img, input_size, interpolation=2)
    
    mean = np.array([0.485, 0.456, 0.406])[np.newaxis, np.newaxis, :]
    std = np.array([0.229, 0.224, 0.225])[np.newaxis, np.newaxis, :]
    img = img / 255
    img -= mean
    img /= std
    img = img.astype(np.float32, copy=False)

    img = img.transpose((2, 0, 1))
    img = img[np.newaxis, :]
    input['image'] = img
    return input

def draw_box(img, res, threshold=0.35):
    """draw box

    Args:
        img_path (str): image path
        res (numpy): output
        threshold (float, optional): . Defaults to 0.2.
    """
    # img = cv2.imread(img_path)
    label_list = ['Boat','People','bottle','Chair','Table','Cat','Dog','Motorbike','Bicycle','Cup','Car','Bus']
    for i in res:
        label = int(i[0])
        score = i[1]
        if score < threshold:
            continue
        print("Something exists! It is ", end="")
        print(label_list[label])
        xmin, ymin, xmax, ymax = i[2:]
        cv2.rectangle(img, (int(xmin), int(ymin)), 
            (int(xmax), int(ymax)), (0, 255, 0))
    return img

class Example(QWidget):

    def __init__(self,config):
        super().__init__()
        self.config = config

        self.img_width, self.img_height = [120 * x for x in [4, 3]]

        self.initCamera()
        self.initUI()
        self.initModel()

    def initCamera(self):
        # 开启视频通道
        self.camera_id = 0 # 为0时表示视频流来自摄像头
        self.camera = cv2.VideoCapture()  # 视频流
        self.camera.open(self.camera_id)

        # 通过定时器读取数据
        self.flush_clock = QTimer()  # 定义定时器，用于控制显示视频的帧率
        self.flush_clock.start(60)   # 定时器开始计时30ms，结果是每过30ms从摄像头中取一帧显示
        self.flush_clock.timeout.connect(self.show_frame)  # 若定时器结束，show_frame()

    def initModel(self):
        # Step 2: initialize ppnc predictor
        self.predictor = PPNCPredictor(self.config)
        self.predictor.load()

    def initUI(self):
        grid = QGridLayout()
        self.setLayout(grid)

        self.Img_Box = QLabel()  # 定义显示视频的Label
        self.Img_Box.setFixedSize(self.img_width, self.img_height)
        grid.addWidget(self.Img_Box, 0, 0, 20, 20)

        self.setWindowTitle('test')
        self.show()

    def show_frame(self):
        # self.player.play()
        _, img = self.camera.read()  # 从视频流中读取
        # print(img.shape)

        try:
            feeds = preprocess(img)
            self.predictor.set_inputs(feeds)
            self.predictor.run()
            res = self.predictor.get_output(0)
            img = draw_box(img, res)
        except:
            print('infer error')

        img = cv2.resize(img, (self.img_width, self.img_height))  # 把读到的帧的大小重新设置为 640x480
        showImage = QImage(img, img.shape[1], img.shape[0], QImage.Format_RGB888)
        self.Img_Box.setPixmap(QPixmap.fromImage(showImage))

if __name__ == '__main__':
    # Step 1: parse config file
    assert len(sys.argv) >= 3, "config and image file must be provided"
    config_file = sys.argv[1]
    image_file = sys.argv[2]
    config = parse_config(config_file)

    app = QApplication(sys.argv)
    ex = Example(config)
    sys.exit(app.exec_())

最后简单展示一下部署效果：

检测环境	检测结果

检测环境看着可能比较亮，但实际上很暗，除了屏幕和开发板的指示灯外，没有太多光源。可以看到确实能够检测到水杯的出现~

请点击此处查看本环境基本用法.

Please click here for more detailed instructions.

此文章为搬运
原项目链接