English | [简体中文](README.md) # PP-Human— a Real-Time Pedestrian Analysis Tool 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 action recognition. 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 achieves real-time analysis on the T4 server. Community intelligent management supportted by PP-Human, please refer to this [AI Studio project](https://aistudio.baidu.com/aistudio/projectdetail/3679564) for quick start tutorial. Full-process operation tutorial of PP-Human, covering training, deployment, action expansion, please refer to this [AI Studio project](https://aistudio.baidu.com/aistudio/projectdetail/3842982). ## I. Environment Preparation Requirement: PaddleDetection version >= release/2.4 or develop The installation of PaddlePaddle and PaddleDetection ``` # PaddlePaddle CUDA10.1 python -m pip install paddlepaddle-gpu==2.2.2.post101 -f https://www.paddlepaddle.org.cn/whl/linux/mkl/avx/stable.html # PaddlePaddle CPU python -m pip install paddlepaddle -i https://mirror.baidu.com/pypi/simple # Clone the PaddleDetection repository cd git clone https://github.com/PaddlePaddle/PaddleDetection.git # Install other dependencies cd PaddleDetection pip install -r requirements.txt ``` 1. For details of the installation, please refer to this [document](../../docs/tutorials/INSTALL.md) 2. Please install `Paddle-TensorRT` if your want speedup inference by TensorRT. You can download the whl package from [Paddle-whl-list](https://paddleinference.paddlepaddle.org.cn/v2.2/user_guides/download_lib.html#python), or prepare the envs by yourself follows the [Install-Guide](https://www.paddlepaddle.org.cn/inference/master/optimize/paddle_trt.html). ## 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 Weights |Model Inference and Deployment | | :---------: |:---------: |:--------------- | :-------: | :------: | :------: | | Object Detection | Image/Video Input | mAP: 56.3 | 28.0ms |[Link](https://bj.bcebos.com/v1/paddledet/models/pipeline/mot_ppyoloe_l_36e_pipeline.pdparams) |[Link](https://bj.bcebos.com/v1/paddledet/models/pipeline/mot_ppyoloe_l_36e_pipeline.zip) | | Object Tracking | Image/Video Input | MOTA: 72.0 | 33.1ms |[Link](https://bj.bcebos.com/v1/paddledet/models/pipeline/mot_ppyoloe_l_36e_pipeline.pdparams) |[Link](https://bj.bcebos.com/v1/paddledet/models/pipeline/mot_ppyoloe_l_36e_pipeline.zip) | | Attribute Recognition | Image/Video Input Attribute Recognition | mA: 94.86 | 2ms per person | - |[Link](https://bj.bcebos.com/v1/paddledet/models/pipeline/strongbaseline_r50_30e_pa100k.zip) | | Keypoint Detection | Video Input Action Recognition | AP: 87.1 | 2.9ms per person | [Link](https://bj.bcebos.com/v1/paddledet/models/pipeline/dark_hrnet_w32_256x192.pdparams) |[Link](https://bj.bcebos.com/v1/paddledet/models/pipeline/dark_hrnet_w32_256x192.zip) | Action Recognition | Video Input Action Recognition | Precision 96.43 | 2.7ms per person | - |[Link](https://bj.bcebos.com/v1/paddledet/models/pipeline/STGCN.zip) | | ReID | Multi-Target Multi-Camera Tracking | mAP: 98.8 | 1.5ms per person | - |[Link](https://bj.bcebos.com/v1/paddledet/models/pipeline/reid_model.zip) | 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. - The precision on ReID model is evaluated on Market1501. - The inference speed is tested on T4, using TensorRT FP16. The pipeline of preprocess, prediction and postprocess is included. ### 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. Their correspondence is as follows: | Input | Function | Task Type | Config | |-------|-------|----------|-----| | Image | Attribute Recognition | Object Detection Attribute Recognition | DET ATTR | | Single-Camera Video | Attribute Recognition | Multi-Object Tracking Attribute Recognition | MOT ATTR | | Single-Camera Video | Behavior Recognition | Multi-Object Tracking Keypoint Detection Action 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: ``` crop_thresh: 0.5 attr_thresh: 0.5 visual: True MOT: model_dir: output_inference/mot_ppyoloe_l_36e_pipeline/ tracker_config: deploy/pphuman/config/tracker_config.yml batch_size: 1 ATTR: model_dir: output_inference/strongbaseline_r50_30e_pa100k/ batch_size: 8 ``` **Note: ** - For different tasks, users could add `--enable_attr=True` or `--enable_action=True` in command line and do not need to set config file. - if only need to change the model path, users could add `--model_dir det=ppyoloe/` in command line and do not need to set config file. For details info please refer to doc below. ### 3. Inference and Deployment ``` # Pedestrian detection. Specify the config file path and test images python deploy/pphuman/pipeline.py --config deploy/pphuman/config/infer_cfg.yml --image_file=test_image.jpg --device=gpu [--run_mode trt_fp16] # Pedestrian tracking. Specify the config file path and test videos python deploy/pphuman/pipeline.py --config deploy/pphuman/config/infer_cfg.yml --video_file=test_video.mp4 --device=gpu [--run_mode trt_fp16] # Pedestrian tracking. Specify the config file path, the model path and test videos # The model path specified on the command line prioritizes over the config file python deploy/pphuman/pipeline.py --config deploy/pphuman/config/infer_cfg.yml --video_file=test_video.mp4 --device=gpu --model_dir det=ppyoloe/ [--run_mode trt_fp16] # Attribute recognition. Specify the config file path and test videos python deploy/pphuman/pipeline.py --config deploy/pphuman/config/infer_cfg.yml --video_file=test_video.mp4 --device=gpu --enable_attr=True [--run_mode trt_fp16] # Action Recognition. Specify the config file path and test videos python deploy/pphuman/pipeline.py --config deploy/pphuman/config/infer_cfg.yml --video_file=test_video.mp4 --device=gpu --enable_action=True [--run_mode trt_fp16] # Pedestrian Multi-Target Multi-Camera tracking. Specify the config file path and the directory of test videos python deploy/pphuman/pipeline.py --config deploy/pphuman/config/infer_cfg.yml --video_dir=mtmct_dir/ --device=gpu [--run_mode trt_fp16] ``` Other usage please refer to [sub-task docs](./docs) ### 3.1 Description of Parameters | Parameter | Optional or not| Meaning | |-------|-------|----------| | --config | Yes | Config file path | | --model_dir | Option | the model paths of different tasks in PP-Human, with a priority higher than config files. For example, `--model_dir det=better_det/ attr=better_attr/` | | --image_file | Option | Images to-be-predicted | | --image_dir | Option | The path of folders of to-be-predicted images | | --video_file | Option | Videos to-be-predicted | | --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 action 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. | ## III. Introduction to the Solution The overall solution of PP-Human is as follows:
### 1. Object Detection - Use PP-YOLOE L as the model of object detection - For details, please refer to [PP-YOLOE](../../configs/ppyoloe/) and [Detection and Tracking](docs/mot_en.md) ### 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) and [Detection and Tracking](docs/mot_en.md) ### 3. Multi-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 [Multi-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. Action 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 [Action Recognition](docs/action_en.md)