{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 1. PP-Human简介\n",
"PaddleDetection深入探索核心行业的高频场景,提供了行人场景的开箱即用分析工具,支持图片/单镜头视频/多镜头视频/在线视频流多种输入方式,广泛应用于智慧交通、智慧城市、工业巡检等领域。支持服务器端部署及TensorRT加速,T4服务器上可达到实时!\n",
"PP-Human支持四大产业级功能:五大异常行为识别、26种人体属性分析、实时人流计数、跨镜头(ReID)跟踪。\n",
"\n",
"PP-Human由飞桨官方出品,是基于PaddleDetection的行人分析pipeline。\n",
"更多关于PaddleDetection可以点击https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.5/deploy/pipeline 进行了解。\n",
"\n"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 2. 模型效果及应用场景\n",
"### 2.1 PP-Human模型效果:\n",
"\n",
"| 任务 | 端到端速度(ms) | 模型方案 | 模型体积 |\n",
"|:---------:|:---------:|:---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------:|:-------------------------------------------:|\n",
"| 行人检测(高精度) | 25.1ms | [多目标跟踪](https://bj.bcebos.com/v1/paddledet/models/pipeline/mot_ppyoloe_l_36e_pipeline.zip) | 182M |\n",
"| 行人检测(轻量级) | 16.2ms | [多目标跟踪](https://bj.bcebos.com/v1/paddledet/models/pipeline/mot_ppyoloe_s_36e_pipeline.zip) | 27M |\n",
"| 行人跟踪(高精度) | 31.8ms | [多目标跟踪](https://bj.bcebos.com/v1/paddledet/models/pipeline/mot_ppyoloe_l_36e_pipeline.zip) | 182M |\n",
"| 行人跟踪(轻量级) | 21.0ms | [多目标跟踪](https://bj.bcebos.com/v1/paddledet/models/pipeline/mot_ppyoloe_s_36e_pipeline.zip) | 27M |\n",
"| 跨镜跟踪(REID) | 单人1.5ms | [REID](https://bj.bcebos.com/v1/paddledet/models/pipeline/reid_model.zip) | REID:92M |\n",
"| 属性识别(高精度) | 单人8.5ms | [目标检测](https://bj.bcebos.com/v1/paddledet/models/pipeline/mot_ppyoloe_l_36e_pipeline.zip)
[属性识别](https://bj.bcebos.com/v1/paddledet/models/pipeline/strongbaseline_r50_30e_pa100k.zip) | 目标检测:182M
属性识别:86M |\n",
"| 属性识别(轻量级) | 单人7.1ms | [目标检测](https://bj.bcebos.com/v1/paddledet/models/pipeline/mot_ppyoloe_l_36e_pipeline.zip)
[属性识别](https://bj.bcebos.com/v1/paddledet/models/pipeline/strongbaseline_r50_30e_pa100k.zip) | 目标检测:182M
属性识别:86M |\n",
"| 摔倒识别 | 单人10ms | [多目标跟踪](https://bj.bcebos.com/v1/paddledet/models/pipeline/mot_ppyoloe_l_36e_pipeline.zip)
[关键点检测](https://bj.bcebos.com/v1/paddledet/models/pipeline/dark_hrnet_w32_256x192.zip)
[基于关键点行为识别](https://bj.bcebos.com/v1/paddledet/models/pipeline/STGCN.zip) | 多目标跟踪:182M
关键点检测:101M
基于关键点行为识别:21.8M |\n",
"| 闯入识别 | 31.8ms | [多目标跟踪](https://bj.bcebos.com/v1/paddledet/models/pipeline/mot_ppyoloe_l_36e_pipeline.zip) | 182M |\n",
"| 打架识别 | 19.7ms | [视频分类](https://bj.bcebos.com/v1/paddledet/models/pipeline/mot_ppyoloe_l_36e_pipeline.zip) | 90M |\n",
"| 抽烟识别 | 单人15.1ms | [目标检测](https://bj.bcebos.com/v1/paddledet/models/pipeline/mot_ppyoloe_l_36e_pipeline.zip)
[基于人体id的目标检测](https://bj.bcebos.com/v1/paddledet/models/pipeline/ppyoloe_crn_s_80e_smoking_visdrone.zip) | 目标检测:182M
基于人体id的目标检测:27M |\n",
"| 打电话识别 | 单人ms | [目标检测](https://bj.bcebos.com/v1/paddledet/models/pipeline/mot_ppyoloe_l_36e_pipeline.zip)
[基于人体id的图像分类](https://bj.bcebos.com/v1/paddledet/models/pipeline/PPHGNet_tiny_calling_halfbody.zip) | 目标检测:182M
基于人体id的图像分类:45M |\n",
"\n",
"\n",
"\n",
"### 2.2 应用场景:\n",
"| 功能 | 方案优势 | 💡示例图 |\n",
"| --------------------- | ------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------- |\n",
"| **跨镜跟踪(ReID)** | 超强性能:针对目标遮挡、完整度、模糊度等难点特殊优化,实现mAP 98.8、1.5ms/人 | |\n",
"| **属性分析** | 兼容多种数据格式:支持图片、视频、在线视频流输入
高性能:融合开源数据集与企业真实数据进行训练,实现mAP 95.4、2ms/人
支持26种属性:性别、年龄、眼镜、上衣、鞋子、帽子、背包等26种高频属性 | |\n",
"| **行为识别** | 功能丰富:支持摔倒、打架、抽烟、打电话、人员闯入五种高频异常行为识别
鲁棒性强:对光照、视角、背景环境无限制
性能高:与视频识别技术相比,模型计算量大幅降低,支持本地化与服务化快速部署
训练速度快:仅需15分钟即可产出高精度行为识别模型 | |\n",
"| **人流量计数**
**轨迹记录** | 简洁易用:单个参数即可开启人流量计数与轨迹记录功能 | |\n",
"\n",
"\n"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 3. 模型如何使用\n",
"\n",
"(在Jupyter Notebook上运行时需要加\"!\",若是cd命令则需加\"%\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# 克隆PaddleDetection仓库\n",
"%mkdir -p ~/work\n",
"%cd ~/work/\n",
"!git clone https://github.com/PaddlePaddle/PaddleDetection.git\n",
"\n",
"# 安装其他依赖\n",
"%cd PaddleDetection\n",
"%mkdir -p demo_input demo_output\n",
"!pip install -r requirements.txt"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"\n",
"\n",
"### 3.1 配置文件说明\n",
"\n",
"PP-Human相关配置位于```deploy/pipeline/config/infer_cfg_pphuman.yml```中,存放模型路径,该配置文件中包含了目前PP-Human支持的所有功能。如果想要查看某个单一功能的配置,请参见```deploy/pipeline/config/examples/```中相关配置。此外,配置文件中的内容可以通过```-o```命令行参数修改,如修改属性的模型目录,则可通过```-o ATTR.model_dir=\"DIR_PATH\"```进行设置。\n",
"\n",
"功能及任务类型对应表单如下:\n",
"\n",
"| 输入类型 | 功能 | 任务类型 | 配置项 |\n",
"|-------|-------|----------|-----|\n",
"| 图片 | 属性识别 | 目标检测 属性识别 | DET ATTR |\n",
"| 单镜头视频 | 属性识别 | 多目标跟踪 属性识别 | MOT ATTR |\n",
"| 单镜头视频 | 行为识别 | 多目标跟踪 关键点检测 摔倒识别 | MOT KPT SKELETON_ACTION |\n",
"\n",
"例如基于视频输入的属性识别,任务类型包含多目标跟踪和属性识别,具体配置如下:\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"```\n",
"crop_thresh: 0.5\n",
"attr_thresh: 0.5\n",
"visual: True\n",
"\n",
"MOT:\n",
" model_dir: https://bj.bcebos.com/v1/paddledet/models/pipeline/mot_ppyoloe_l_36e_pipeline.zip\n",
" tracker_config: deploy/pipeline/config/tracker_config.yml\n",
" batch_size: 1\n",
" enable: True\n",
"\n",
"ATTR:\n",
" model_dir: https://bj.bcebos.com/v1/paddledet/models/pipeline/PPLCNet_x1_0_person_attribute_945_infer.zip\n",
" batch_size: 8\n",
" enable: True\n",
"```"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"**注意:**\n",
"\n",
"- 如果用户需要实现不同任务,可以在配置文件对应enable选项设置为True。\n",
"\n",
"### 3.2 预测部署"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"#直接使用默认配置或者examples中配置文件,或者直接在`infer_cfg_pphuman.yml`中修改配置:\n",
"\n",
"# 例:行人检测,指定配置文件路径和测试图片,图片输入默认打开检测模型\n",
"!wget -P demo_input -N https://paddledet.bj.bcebos.com/modelcenter/images/PP-Human/human_attr.jpg \n",
"!python deploy/pipeline/pipeline.py --config deploy/pipeline/config/infer_cfg_pphuman.yml --image_file=demo_input/human_attr.jpg --device=gpu --output_dir=demo_output"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# 例:行人属性识别,直接使用examples中配置\n",
"!wget -P demo_input -N https://paddledet.bj.bcebos.com/modelcenter/images/PP-Human/human_attr.mp4\n",
"!python deploy/pipeline/pipeline.py --config deploy/pipeline/config/examples/infer_cfg_human_attr.yml --video_file=demo_input/human_attr.mp4 --device=gpu --output_dir=demo_output"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"#使用命令行进行功能开启,或者模型路径修改:\n",
"\n",
"# 例:行人跟踪,指定配置文件路径,模型路径和测试视频, 命令行中指定的模型路径优先级高于配置文件\n",
"!wget -P demo_input -N https://paddledet.bj.bcebos.com/modelcenter/images/PP-Human/human_count.mp4\n",
"!python deploy/pipeline/pipeline.py --config deploy/pipeline/config/infer_cfg_pphuman.yml -o MOT.enable=True --video_file=demo_input/human_count.mp4 --device=gpu --output_dir=demo_output"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# 例:行为识别,以摔倒识别为例,命令行中开启SKELETON_ACTION模型\n",
"!wget -P demo_input -N https://paddledet.bj.bcebos.com/modelcenter/images/PP-Human/human_falldown.mp4\n",
"!python deploy/pipeline/pipeline.py --config deploy/pipeline/config/infer_cfg_pphuman.yml -o SKELETON_ACTION.enable=True --video_file=demo_input/human_falldown.mp4 --device=gpu --output_dir=demo_output"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"#rtsp推拉流\n",
"\n",
"#对rtsp拉流的支持,使用--rtsp RTSP [RTSP ...]参数指定一路或者多路rtsp视频流,如果是多路地址中间用空格隔开。(或者video_file后面的视频地址直接更换为rtsp流地址),示例如下:\n",
"\n",
"# 例:行人属性识别,单路视频流\n",
"!python deploy/pipeline/pipeline.py --config deploy/pipeline/config/examples/infer_cfg_human_attr.yml -o visual=False --rtsp rtsp://[YOUR_RTSP_SITE] --device=gpu"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# 例:行人属性识别,多路视频流\n",
"!python deploy/pipeline/pipeline.py --config deploy/pipeline/config/examples/infer_cfg_human_attr.yml -o visual=False --rtsp rtsp://[YOUR_RTSP_SITE1] rtsp://[YOUR_RTSP_SITE2] --device=gpu"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# 视频结果推流rtsp\n",
"\n",
"# 预测结果进行rtsp推流,使用--pushurl rtsp:[IP] 推流到IP地址端,PC端可以使用[VLC播放器](https://vlc.onl/)打开网络流进行播放,播放地址为 `rtsp:[IP]/videoname`。其中`videoname`是预测的视频文件名,如果视频来源是本地摄像头则`videoname`默认为`output`.\n",
"\n",
"# 例:行人属性识别,单路视频流,该示例播放地址为 rtsp://[YOUR_SERVER_IP]:8554/test_video\n",
"!python deploy/pipeline/pipeline.py --config deploy/pipeline/config/examples/infer_cfg_human_attr.yml --video_file=test_video.mp4 --device=gpu --pushurl rtsp://[YOUR_SERVER_IP]:8554\n",
"\n",
"# 注:\n",
"# 1. rtsp推流服务基于 [rtsp-simple-server](https://github.com/aler9/rtsp-simple-server), 如使用推流功能请先开启该服务.\n",
"# 2. rtsp推流如果模型处理速度跟不上会出现很明显的卡顿现象,建议跟踪模型使用ppyoloe_s版本,即修改配置中跟踪模型mot_ppyoloe_l_36e_pipeline.zip替换为mot_ppyoloe_s_36e_pipeline.zip。"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"\n",
"### 3.3 Jetson部署说明\n",
"\n",
"由于Jetson平台算力相比服务器有较大差距,有如下使用建议:\n",
"\n",
"1. 模型选择轻量级版本,特别是跟踪模型,推荐使用`ppyoloe_s: https://bj.bcebos.com/v1/paddledet/models/pipeline/mot_ppyoloe_s_36e_pipeline.zip`\n",
"2. 开启跟踪跳帧功能,推荐使用2或者3. `skip_frame_num: 3`\n",
"\n",
"使用该推荐配置,在TX2平台上可以达到较高速率,经测试属性案例达到20fps。\n",
"\n",
"可以直接修改配置文件(推荐),也可以在命令行中修改(字段较长,不推荐)。\n",
"\n",
"\n",
"### 参数说明\n",
"\n",
"| 参数 | 是否必须|含义 |\n",
"|-------|-------|----------|\n",
"| --config | Yes | 配置文件路径 |\n",
"| -o | Option | 覆盖配置文件中对应的配置 |\n",
"| --image_file | Option | 需要预测的图片 |\n",
"| --image_dir | Option | 要预测的图片文件夹路径 |\n",
"| --video_file | Option | 需要预测的视频,或者rtsp流地址 |\n",
"| --rtsp | Option | rtsp视频流地址,支持一路或者多路同时输入 |\n",
"| --camera_id | Option | 用来预测的摄像头ID,默认为-1(表示不使用摄像头预测,可设置为:0 - (摄像头数目-1) ),预测过程中在可视化界面按`q`退出输出预测结果到:output/output.mp4|\n",
"| --device | Option | 运行时的设备,可选择`CPU/GPU/XPU`,默认为`CPU`|\n",
"| --pushurl | Option| 对预测结果视频进行推流的地址,以rtsp://开头,该选项优先级高于视频结果本地存储,打开时不再另外存储本地预测结果视频, 默认为空,表示没有开启|\n",
"| --output_dir | Option|可视化结果保存的根目录,默认为output/|\n",
"| --run_mode | Option |使用GPU时,默认为paddle, 可选(paddle/trt_fp32/trt_fp16/trt_int8)|\n",
"| --enable_mkldnn | Option | CPU预测中是否开启MKLDNN加速,默认为False |\n",
"| --cpu_threads | Option| 设置cpu线程数,默认为1 |\n",
"| --trt_calib_mode | Option| TensorRT是否使用校准功能,默认为False。使用TensorRT的int8功能时,需设置为True,使用PaddleSlim量化后的模型时需要设置为False |\n",
"| --do_entrance_counting | Option | 是否统计出入口流量,默认为False |\n",
"| --draw_center_traj | Option | 是否绘制跟踪轨迹,默认为False |\n",
"| --region_type | Option | 'horizontal'(默认值)、'vertical':表示流量统计方向选择;'custom':表示设置车辆禁停区域 |\n",
"| --region_polygon | Option | 设置禁停区域多边形多点的坐标,无默认值 |\n",
"| --illegal_parking_time | Option | 设置禁停时间阈值,单位秒(s),-1(默认值)表示不做检查 |\n"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 4. 方案介绍\n",
"PP-Human 整体方案如下图所示:\n",
"\n",
"