Steps to Train a YOLO model

This document describes the process for training a YOLO object detection model to identify cones used in Formula Trinity's perception module. The model is trained on the FSOCO dataset, a standardized dataset for Formula Student driverless teams.

---

1. Dataset Setup

Download Dataset

Access the FSOCO training/validation dataset used for YOLOv5 training from the shared Google Drive:

🔗 FSOCO Dataset (TCD access required)

The above dataset maintains a 80/20 training–validation split.

dataset/
├── train/
│   ├── images/
│   └── labels/
└── val/
    ├── images/
    └── labels/

---

2. Environment Setup

Install Ultralytics

YOLOv12 is part of the unified Ultralytics framework.

pip install ultralytics

If the yolo executable is not found in your PATH, update it:

export PATH=$PATH:/home/$USER/.local/bin

Verify installation:

yolo

---

3. Create the Dataset Configuration File

Create a file named dataset.yml to describe your dataset layout and class structure.

path: /home/$USER/fsoco  # Update path to the dataset root

train: train/images
val: val/images

nc: 6
names:
  0: yellow_cone
  1: orange_cone
  2: large_orange_cone
  3: blue_cone
  4: other_cone
  5: unknown_cone

- YOLO uses this YAML file to locate your training/validation images and to define the class labels for object detection.

---

4. Model Selection & Rationale

Why use the YOLO-s (small) model variant?

• Speed-critical environment: The perception stack in Formula Trinity runs in real-time on a Jetson AGX Orin.
• Trade-off balance: The s model provides a strong balance between inference speed and accuracy.

---

5. Train the Model

• Create a bash script to automate training. E.g. train_v12.sh.
• Note: You may need to download the base model from the ultralytics repo and place in /home/$USER/ft/. See here

#!/usr/bin/env bash

yolo detect train \
  data=/home/$USER/ft/dataset.yml \
  model=/home/$USER/ft/yolo12s.pt \    
  epochs=50 \
  imgsz=640 \
  batch=10

Make the script executable:

chmod 700 train_v12.sh

Execute training:

./train_v12.sh

Training typically takes ~2 hours for 50 epochs (on fta-pc).

---

6. Outputs

After training completes, the results are saved under:

runs/detect/train/

Key files:

• weights/best.pt → best-performing model weights
• results.png → training curve visualization

---

7. Validate the Model

To validate the trained model on the validation set:

yolo detect val model=/home/$USER/ft/runs/detect/train/weights/best.pt data=/home/$USER/ft/dataset.yml

This will output accuracy metrics including precision, recall, and mAP, along with visualized validation samples.

---

8. Next Steps

Once validated, the model (best.pt) can be integrated into the perception stack for cone detection.

---