GOFFIN Gaspard, LOMBARDO Théo, NZITAT William

Person Detection, Tracking, and Re-Identification Toolbox

Overview

This project is a modular toolbox designed for detecting, tracking, and re-identifying people in various environments. It combines state-of-the-art detection algorithms, tracking methods, and re-identification (REID) modules to enable robust person identification in real-world applications.

Features

• Detection: Supports YOLO Classic, YOLO Ultralytics, and Ground Truth-based detection.
• Tracking: Implements SORT, DeepSORT, and Centroid-based tracking.
• Re-Identification: Integrates Facenet and TorchReid for re-identification using pre-trained .pkl files.
• Modularity: Designed to allow seamless integration of different detectors, trackers, and REID modules.
• Multithreading: Capable of processing multiple streams simultaneously for efficient performance.

---

Configurations

Detectors

Parameter	YOLO Classic	YOLO Ultralytics	GT
dt_name	YOLO_Classic	YOLO_Ultralytics	GT
nms	0.45	-	-
conf	0.5	0.5	-
class_file	Path to class file (e.g., coco.names)	-	-
model_cfg_file	Path to YOLO configuration file (e.g., yolov4.cfg)	-	-
model_weights	Path to YOLO weights file (e.g., yolov4.weights)	-	-
model_image_size	416	416	-
repspoint_calibration	0.25	0.25	-
iou	-	0.7	-
imgsz	-	416	-
max_det	-	100	-
gt_file	-	-	Path to ground truth file
gt_map_file	-	-	Path to ground truth map file

---

Trackers

Parameter	Centroid	SORT	DeepSORT
tk_name	Centroid	SORT	DeepSORT
max_spread	64	-	-
max_age	-	1	-
min_hits	-	3	-
iou_threshold	-	0.3	-
nn_budget	-	-	100
nms_max_overlap	-	-	0.5
max_cosine_distance	-	-	0.1
model_file	-	-	Path to DeepSORT model file

---

Re-Identification (REID)

Parameter	Facenet	TorchReid
ri_name	Facenet	Torchreid
gpu_mem	0.585	-
model_det	Path to Facenet detection model	-
model_file	Path to Facenet model file	Path to TorchReid model file
classifier_pkl	Path to classifier .pkl file	Path to TorchReid classifier .pkl file
train_data	Path to training dataset	Path to training dataset
batch_size	1000	-
min_confidence	0.75	0.35
yl_h_calibration	[-125, 75]	-
yl_w_calibration	[-55, 55]	-
device	-	cpu or gpu

---

Main Configuration

Set the main modules in the main_config.yaml file:

# Main config:
###########################################################
# detector: None | YOLO_Classic | YOLO_Ultralytics
# tracker: None | Centroid | SORT | DeepSORT
# reider: None | Facenet | Torchreid
###########################################################
detector: ...
tracker: ...
reider: ...

Project Structure

├── project/
│   ├── cfg/  # Basic configuration  
│   ├── my_config/ # Project configurations
│   ├── data/ # Store videos there
│   ├── comparebox.py/  
│   ├── toolbox.py/
│ 
└── README.md

Setup

1. Clone the repository:

    git clone https://github.com/DonWilliams09/pyppbox_project.git
    cd project
   

2. Place the .pkl files provided by the first team in the appropriate REID module directory.
3. Configure paths in the examples/configs directory to match your setup.
4. Add the video you want to process in the data file

Workflow

The first team focuses on registering new people and training a ReID module. These trained models are saved in .pkl files and provided to the tracking team. The tracking team uses these .pkl files to perform the tracking and re-identification tasks, integrating them into the pipeline with detectors and trackers.

Usage

1. Toolbox File

To run the detection, tracking, and re-identification process, use the following command:

python toolbox.py

In toolbox.py, you can specify the the configuration used by the detectors and trackers.

setConfigDir(config_dir="your_config", load_all=True)

2. Compare Toolbox

To compare the detection, tracking, and re-identification processes using two different re-ids use the following command:

python comparebox.py

In toolbox.py, you can specify the paths for the configuration directories used by the detectors and trackers.

config_dir_one = "/path/to/config_one"
config_dir_two = "/path/to/config_two"
ppb_task(config_dir_one, config_dir_two)

3. Input File

You can run the toolbox either on a pre-saved video

input_video = "path_to_the_video"
cap = cv2.VideoCapture(input_video)

or in real-time using a connected camera.

cap = cv2.VideoCapture(0)

The output will display the processed video stream, where detected and re-identified people will be labeled with their IDs. If the toolbox fails to recognize a person, they will be required to register for inclusion in the ReID module's training. Once completed, the .pkl file will be updated in the toolbox, enabling the person to be identified in future detections.