Transport YOLO Robot — Pipeline de vision par ordinateur

Python pipeline for data preparation, training, and object detection for a robotic transportation system.

Transport YOLO Robot is a computer vision project that allows merging multiple YOLO format datasets and training custom detectors. It recognizes road signs, traffic lights, and the additional class Bart Simpson, while keeping the weights, metrics, and training graphs.

01

Problem addressed

The necessary data for detection is spread across multiple datasets using different classes and identifiers. The project must standardize them in order to train a single detector capable of recognizing multiple road elements and an additional visual class.

02

Objectives

Merging multiple YOLO format datasets.
Standardizing class identifiers in annotations.
Creating a three-class dataset.
Creating a second six-class dataset.
Training custom detectors using transfer learning.
Producing best.pt and last.pt weights.
Generating validation metrics and graphs.
Performing local object and QR code detection.
Producing voice output from detected content.

03

Technical solution

The project uses two Python scripts to merge images and annotations from multiple datasets. fusionner_datasets.py generates dataset_final with the classes stop, vitesse_80, and feux_rouge. fusionner_v2.py generates dataset_v2 with the classes stop, vitesse_80, feux_rouge, feux_vert, feux_orange, and bart_simpson. The scripts entrainer.py and entrainer_v2.py then train two custom detectors for 50 and 80 epochs, respectively. The weights, metrics, confusion matrices, and curves are saved in runs/detect/. The local application pc_main.py performs visual inference, detects QR codes with OpenCV, and uses speech synthesis. However, it currently loads the generic model yolov8n.pt and not the custom weights runs/detect/best_2/weights/best.pt.

04

Architecture

YOLO dataset sources
Stop Sign dataset
Vitesse 80 dataset
Red Lights dataset
Green Lights dataset
Orange Lights dataset
Bart Simpson dataset
Python merge scripts
merge_datasets.py script
dataset_final dataset
merge_v2.py script
dataset_v2 dataset
Remapped YOLO annotations
data.yaml files
Initial YOLOv8n model
train.py script
Training for 50 epochs
robot_panneaux-4 run
train_v2.py script
Training for 80 epochs
best_2 run
best.pt weights
last.pt weights
results.csv file
Confusion matrices
Precision curves
Recall curves
mAP curves
Local application pc_main.py
Video inference
Object detection
QR code detection
Speech synthesis
Audio playback

05

Documented results

dataset_final contains 726 image and annotation pairs.
dataset_final groups the Stop, Speed 80, and Red Lights classes.
dataset_v2 contains 1,046 image and annotation pairs.
dataset_v2 contains 509 Stop images.
dataset_v2 contains 50 Speed 80 images.
dataset_v2 contains 99 Red Lights images.
dataset_v2 contains 200 Green Lights images.
dataset_v2 contains 88 Orange Lights images.
dataset_v2 contains 100 Bart Simpson images.
The best_2 run contains the best.pt and last.pt weights.
At epoch 80, the best_2 run reaches an accuracy of 0.98033.
At epoch 80, the best_2 run reaches a recall of 0.97377.
At epoch 80, the best_2 run reaches an mAP@50 of 0.98464.
At epoch 80, the best_2 run reaches an mAP@50-95 of 0.87617.
The robot_panneaux-4 run contains the best.pt and last.pt weights.
At epoch 50, the robot_panneaux-4 run reaches an accuracy of 0.99771.
At epoch 50, the robot_panneaux-4 run reaches a recall of 0.99942.
At epoch 50, the robot_panneaux-4 run reaches an mAP@50 of 0.99452.
At epoch 50, the robot_panneaux-4 run reaches an mAP@50-95 of 0.88480.
The configured validation sets are identical to the training sets.
The obtained metrics therefore do not correspond to an independent evaluation.
No automated testing is present.
No CI/CD workflow is present.
No database is used.
No local API is implemented.
No versioned dependency file is present.