若该文为原创文章，转载请注明原文出处。

分享一个行人属性分析系统，识别行人，并标记每个人的属性。

项目代码来自公众号渡码的项目。

本人用Win10复现完整项目，并记录过程。

源码会上传到github,可以自行下载测试。

Yinyifeng18/AI_pedestrain_attributes: 行人属性识别 (github.com)

一、介绍

实现行人属性分析系统需要 3 个步骤：

用 YOlOv5 识别行人

用 ByteTrack 跟踪标记同一个人

训练多标签图像分类网络，识别行人 26 个属性

二、环境安装

使用环境是win10下CPU版本

1、安装Anaconda

下载地址：Miniconda — Anaconda documentation

下载安装即可。

2、创建虚拟机

conda create -n pedestrain_env python=3.8

3、激活

conda activate pedestrain_env

三、安装轮子

pip install loguru -i https://pypi.tuna.tsinghua.edu.cn/simple  
pip install lapx -i https://pypi.tuna.tsinghua.edu.cn/simple 
pip install thop -i https://pypi.tuna.tsinghua.edu.cn/simple 

pip install numpy==1.23.5 -i https://pypi.tuna.tsinghua.edu.cn/simple

安装cython_bbox，win下无法直接安装

先安装cython

pip install cython -i https://pypi.tuna.tsinghua.edu.cn/simple

下载cython_bbox-0.1.3

https://pypi.org/project/cython-bbox/0.1.3/#files

修改里面的setup.py文件

extra_compile_args = [-Wno-cpp]为extra_compile_args = {'gcc': ['/Qstd=c99']}

安装

python setup.py build_ext install

安装时如果出Microsoft Visual C++ 14.0 or greater is required. Get it with "Microsoft C++ Build Tools"的错误处理方法安装VS C++

Microsoft Visual C++ 14.0 or greater is required. Get it with "Microsoft C++ Build Tools"的解决办法 - 知乎 (zhihu.com)

四、安装paddlepadddle

电脑无GPU，安装的是CPU版本

python -m pip install paddlepaddle -i https://mirror.baidu.com/pypi/simple

五、安装paddleclas

paddleclas识别的属性包括以下 10 类

1、安装

pip install paddleclas  -i https://mirror.baidu.com/pypi/simple

2、测试

测试数据下载：

https://paddleclas.bj.bcebos.com/data/PULC/pulc_demo_imgs.zip

下载后解压到当前目录

测试有两种方法：

1）、使用命令行快速预测

paddleclas --model_name=person_attribute --infer_imgs=./pulc_demo_imgs/person_attribute/090004.jpg

2）、python测试

创建paddleclas_demo.py文件，内容如下：

import paddleclas
model = paddleclas.PaddleClas(model_name="person_attribute")
result = model.predict(input_data="pulc_demo_imgs/person_attribute/090004.jpg")
print(next(result))

运行测试

测试结果相同。

这里只测试不做训练，如果想训练，参考文章PaddleClas/docs/zh_CN/models/PULC/PULC_person_attribute.md at release/2.5 · PaddlePaddle/PaddleClas (github.com)

六、安装yolov5

1、下载yolov5

git clone https://github.com/ultralytics/yolov5

2、安装

cd yolov5
pip install -r requirements.txt  -i https://pypi.tuna.tsinghua.edu.cn/simple

3、测试

创建yolov5_demo.py文件，内容如下：

import torch

# Model
model = torch.hub.load('yolov5',
                                         'custom',
                                         path='./yolov5/weights/yolov5s.pt',
                                         source='local')# or yolov5n - yolov5x6, custom

# Images
img = "./zidane.jpg"  # or file, Path, PIL, OpenCV, numpy, list

# Inference
results = model(img)

# Results
results.print()  # or .show(), .save(), .crop(), .pandas(), etc.

需要先下载yolov5s.pt文件。

测试结果

七、下载ByteTrack

git clone https://github.com/ifzhang/ByteTrack.git

到此文件准备完成

八、代码分析测试

代码只有三个，全部附上

track.py

# track.py
from dataclasses import dataclass
import numpy as np
import cv2
import paddleclas
from pedestrain_attr_dict import *
import sys
sys.path.append('./ByteTrack/')
from yolox.tracker.byte_tracker import BYTETracker, STrack


@dataclass(frozen=True)
class BYTETrackerArgs:
    track_thresh: float = 0.25
    track_buffer: int = 30
    match_thresh: float = 0.8
    aspect_ratio_thresh: float = 3.0
    min_box_area: float = 1.0
    mot20: bool = False


class Detection(object):
    def __init__(self, ltrb, track_id, person_attr):
        self.track_id = track_id
        self.ltrb = None
        self.sex = ''
        self.age = ''
        self.front = '未知'
        self.has_glasses = '否'
        self.has_hat = '否'
        self.bag = '未知'
        self.upper = ''
        self.lower = ''
        self.boots = '否'

        self.update(ltrb, person_attr)

    def update(self, ltrb, attr):
        self.ltrb = ltrb
        if attr is not None:
            self.sex = '女' if attr[0] == 'Female' else '男'
            self.age = age_dict[attr[1]]
            self.front = direct_list[attr[2]]
            self.has_glasses = '是' if attr[3] == 'Glasses: True' else '否'
            self.has_hat = '是' if attr[4] == 'Hat: True' else '否'
            self.bag = bag_dict[attr[6]]

            # 上半身
            self.upper = ' '.join([upper_dict[up] for up in attr[7].replace('Upper: ', '').split(' ')])
            # 下半身
            self.lower = ' '.join([lower_dict[lo] for lo in attr[8].replace('Lower:  ', '').split(' ')])
            self.boots = '是' if attr[9] == 'Boots' else '否'


class PedestrainTrack(object):
    def __init__(self):
        self.byte_tracker = BYTETracker(BYTETrackerArgs())
        self.detection_dict = {}

        # 行人属性模型
        self.pedestrain_attr_model = paddleclas.PaddleClas(model_name="person_attribute")

    def update_track(self, boxes, frame):
        tracks = self.byte_tracker.update(
            output_results=boxes,
            img_info=frame.shape,
            img_size=frame.shape
        )

        new_detection_dict = {}
        for track in tracks:
            l, t, r, b = track.tlbr.astype(np.int32)
            track_id = track.track_id
            # 调用行人检测模型，识别行人属性
            track_box = frame[t:b, l:r]
            # print(track_box.shape)
            # cv2.imwrite('a.jpg', track_box)
            person_attr_res = self.pedestrain_attr_model.predict(track_box)
            attr = None
            try:
                for i in person_attr_res:
                    attr = i[0]['attributes']
            except:
                pass

            if track_id in self.detection_dict:
                detection = self.detection_dict[track_id]
                detection.update((l, t, r, b), attr)
            else:
                detection = Detection((l, t, r, b), track_id, attr)

            new_detection_dict[track_id] = detection

        self.detection_dict = new_detection_dict

        return self.detection_dict

pedestrain_attr_dict.py

gender_dict = {
    'Female': '女',
    'Male': '男',
}

age_dict = {
    'AgeLess18': '小于18岁',
    'Age18-60': '18-60岁',
    'AgeOver60': '大于60岁'
}

# 朝向
direct_list = {
    'Front': '正面',
    'Side': '侧面',
    'Back': '背面'
}

# 背包
bag_dict = {
    'HandBag': '手提包',
    'ShoulderBag': '单肩包',
    'Backpack': '双肩包',
    'No bag': '未背包'
}

# 上衣风格
upper_dict = {
    'LongSleeve': '长袖',
    'ShortSleeve': '短袖',
    'UpperStride': '带条纹',
    'UpperLogo': '带logo',
    'UpperPlaid': '带格子',
    'UpperSplice': '拼接风格'
}

# 下身风格
lower_dict = {
    'LowerStripe': '带条纹',
    'LowerPattern': '带图案',
    'LongCoat': '长外套',
    'Trousers': '长裤',
    'Shorts': '短裤',
    'Skirt&Dress': '短裙&裙子'
}

pedestrain_attributes.py

import cv2
import torch
import numpy as np
from PIL import Image, ImageDraw, ImageFont
from track import PedestrainTrack


def cv2_add_chinese_text(img, text, position, text_color=(0, 255, 0), text_size=30):
    img = Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
    # 创建一个可以在给定图像上绘图的对象
    draw = ImageDraw.Draw(img)
    # 字体的格式
    fontStyle = ImageFont.truetype("./fonts/simsun.ttc", text_size, encoding="utf-8")
    # 绘制文本
    draw.text(position, text, text_color, font=fontStyle)
    # 转换回OpenCV格式
    return cv2.cvtColor(np.asarray(img), cv2.COLOR_RGB2BGR)


class PedestrainAttrDetection(object):
    def __init__(self):
        self.yolo_model = torch.hub.load('yolov5',
                                         'custom',
                                         path='./yolov5/weights/yolov5s.pt',
                                         source='local')
        self.yolo_model.conf = 0.6
        self.tracker = PedestrainTrack()

    def plot_detection(self, person_track_dict, frame):
        for track_id, detection in person_track_dict.items():
            l, t, r, b = detection.ltrb
            track_id = detection.track_id

            cv2.rectangle(frame, (l, t), (r, b), (0, 255, 0), 1)
            cv2.putText(frame, f'id-{track_id}', (l + 2, t - 3), cv2.FONT_HERSHEY_PLAIN, 1, (0, 0, 255), 1)
            top_margin = 20
            font_size = 14
            font_color = (255, 0, 255)
            frame = cv2_add_chinese_text(frame, f'性别：{detection.sex}', (l+2, t+top_margin), font_color, font_size)
            frame = cv2_add_chinese_text(frame, f'年龄：{detection.age}', (l+2, t+top_margin*2), font_color, font_size)
            frame = cv2_add_chinese_text(frame, f'朝向：{detection.front}', (l+2, t+top_margin*3), font_color, font_size)
            frame = cv2_add_chinese_text(frame, f'戴眼镜：{detection.has_glasses}', (l+2, t+top_margin*4), font_color, font_size)
            frame = cv2_add_chinese_text(frame, f'戴帽子：{detection.has_hat}', (l+2, t+top_margin*5), font_color, font_size)
            frame = cv2_add_chinese_text(frame, f'背包：{detection.bag}', (l+2, t+top_margin*6), font_color, font_size)
            frame = cv2_add_chinese_text(frame, f'上半身：{detection.upper}', (l+2, t+top_margin*7), font_color, font_size)
            frame = cv2_add_chinese_text(frame, f'下半身：{detection.lower}', (l+2, t+top_margin*8), font_color, font_size)
            frame = cv2_add_chinese_text(frame, f'穿靴：{detection.boots}', (l+2, t+top_margin*9), font_color, font_size)
            # cv2.putText(frame, f'sex-{sex}', (l + 2, t + 10), cv2.FONT_HERSHEY_PLAIN, 1, (0, 0, 255), 1)
        return frame

    @staticmethod
    def yolo_pd_to_numpy(yolo_pd):
        box_list = yolo_pd.to_numpy()
        detections = []
        for box in box_list:
            l, t = int(box[0]), int(box[1])
            r, b = int(box[2]), int(box[3])

            conf = box[4]

            detections.append([l, t, r, b, conf])
        return np.array(detections, dtype=float)

    def detect(self, video_file):
        cap = cv2.VideoCapture(video_file)
        video_w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
        video_h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
        fps = round(cap.get(cv2.CAP_PROP_FPS))
        print(fps)
        # video_writer = cv2.VideoWriter('./video_p2.mp4', cv2.VideoWriter_fourcc(*'H264'), fps, (video_w, video_h))

        while True:
            ret, frame = cap.read()
            if not ret or frame is None:
                break

            yolo_det_results = self.yolo_model(frame[:, :, ::-1])
            pd = yolo_det_results.pandas().xyxy[0]
            person_pd = pd[pd['name'] == 'person']

            person_det_boxes = self.yolo_pd_to_numpy(person_pd)
            person_track_dict = self.tracker.update_track(person_det_boxes, frame)
            frame = self.plot_detection(person_track_dict, frame)

            cv2.imshow('pedestrain attributes detect', frame)

            # video_writer.write(frame)
            if cv2.waitKey(10) & 0xFF == ord('q'):
                return


if __name__ == '__main__':
    PedestrainAttrDetection().detect('./video.mp4')

代码不多，比较容易看懂

九、总结
测试过程中由于使用的是CPU安装cython_bbox花了一点时间.环境搭建还是蛮麻烦的。

如有侵权，或需要完整代码，请及时联系博主。