PP-Human serves as the first open-source tool of real-time pedestrian anaylsis relying on the PaddlePaddle deep learning framework. Versatile and efficient in deployment, it has been used in various senarios. PP-Human
offers many input options, including image/single-camera video/multi-camera video, and covers multi-object tracking, attribute recognition, and behavior analysis. PP-Human can be applied to intelligent traffic, the intelligent community, industiral patrol, and so on. It supports server-side deployment and TensorRT acceleration，and even can achieve real-time analysis on the T4 server.
## I. Environment Preparation
Requirement: PaddleDetection version >= release/2.4
The installation of PaddlePaddle and PaddleDetection
For details of the installation, please refer to this [document](docs/tutorials/INSTALL_cn.md)
## II. Quick Start
### 1. Model Download
To make users have access to models of different scenarios, PP-Human provides pre-trained models of object detection, attribute recognition， behavior recognition, and ReID.
| Task | Scenario | Precision | Inference Speed（FPS） | Model Inference and Deployment |
Then, unzip the downloaded model to the folder `./output_inference`.
**Note: **
- The model precision is decided by the fusion of datasets which include open-source datasets and enterprise ones.
- When the inference speed is T4, use TensorRT FP16.
### 2. Preparation of Configuration Files
Configuration files of PP-Human are stored in ```deploy/pphuman/config/infer_cfg.yml```. Different tasks are for different functions, so you need to set the task type beforhand.
| Single-Camera Video | Attribute Recognition | Multi-Object Tracking Attribute Recognition | MOT ATTR |
| Single-Camera Video | Behavior Recognition | Multi-Object Tracking Keypoint Detection Behavior Recognition | MOT KPT ACTION |
For example, for the attribute recognition with the video input, its task types contain multi-object tracking and attribute recognition, and the config is:
| --camera_id | Option | ID of the inference camera is -1 by default (means inference without cameras，and it can be set to 0 - (number of cameras-1)), and during the inference, click `q` on the visual interface to exit and output the inference result to output/output.mp4|
| --enable_attr| Option | Enable attribute recognition or not |
| --enable_action| Option | Enable behavior recognition or not |
| --device | Option | During the operation，available devices are `CPU/GPU/XPU`，and the default is `CPU`|
| --output_dir | Option| The default root directory which stores the visualization result is output/|
| --run_mode | Option | When using GPU，the default one is paddle, and all these are available（paddle/trt_fp32/trt_fp16/trt_int8）.|
| --enable_mkldnn | Option |Enable the MKLDNN acceleration or not in the CPU inference, and the default value is false |
| --cpu_threads | Option| The default CPU thread is 1 |
| --trt_calib_mode | Option| Enable calibration on TensorRT or not, and the default is False. When using the int8 of TensorRT，it should be set to True; When using the model quantized by PaddleSlim, it should be set to False. |
- For details, please refer to [PP-YOLOE](../../configs/ppyoloe/)
### 2. Multi-Object Tracking
- Conduct multi-object tracking with the SDE solution
- Use PP-YOLOE L as the detection model
- Use the Bytetrack solution to track modules
- For details, refer to [Bytetrack](configs/mot/bytetrack)
### 3. Cross-Camera Tracking
- Use PP-YOLOE + Bytetrack to obtain the tracks of single-camera multi-object tracking
- Use ReID（centroid network）to extract features of the detection result of each frame
- Match the features of multi-camera tracks to get the cross-camera tracking result
- For details, please refer to [Cross-Camera Tracking](docs/mtmct_en.md)
### 4. Attribute Recognition
- Use PP-YOLOE + Bytetrack to track humans
- Use StrongBaseline（a multi-class model）to conduct attribute recognition, and the main attributes include age, gender, hats, eyes, clothing, and backpacks.
- For details, please refer to [Attribute Recognition](docs/attribute_en.md)
### 5. Behavior Recognition
- Use PP-YOLOE + Bytetrack to track humans
- Use HRNet for keypoint detection and get the information of the 17 key points in the human body
- According to the changes of the key points of the same person within 50 frames, judge whether the action made by the person within 50 frames is a fall with the help of ST-GCN
- For details, please refer to [Behavior Recognition](docs/action_en.md)
