🎯 Advanced Usage — Custom Objects

This section explains how to change the robot's object detection targets by creating custom datasets, training your own YOLO model, and optimizing detection performance for your LIMO system.

---

🛠 Changing Object Detection Classes

Your YOLO detector (object_detector.py) currently loads a custom model:

self.model = YOLO("/home/agilex/krish_ws/runs/detect/mycobot_final2/weights/best.pt")

To change classes:

1. Train a new YOLO model with your desired objects.
2. Replace the .pt file path in the detector node.
3. Ensure self.model.names contains the correct class labels.
4. Restart the system to load the new model.

💡 Note: If you only want to filter detections to certain classes, you can check cls_name inside the YOLO callback and ignore others.

---

📸 Custom Dataset Creation

1. Collect Images

Use the LIMO's camera:

ros2 run rqt_image_view rqt_image_view

Take screenshots or record video, then extract frames.

2. Annotate

Use LabelImg or Label Studio:

labelImg /path/to/images classes.txt

Export format: YOLO TXT format (one file per image).

3. Organize Dataset

dataset/
 ├── images/
 │    ├── train/
 │    ├── val/
 ├── labels/
 │    ├── train/
 │    ├── val/
 └── data.yaml

Example data.yaml:

train: /absolute/path/to/dataset/images/train
val: /absolute/path/to/dataset/images/val
nc: 3
names: ["objectA", "objectB", "objectC"]

---

🧠 Model Training Procedures

Train in Python

from ultralytics import YOLO

model = YOLO("yolov8n.pt")  # or yolov8s.pt for more accuracy
model.train(
    data="/path/to/data.yaml",
    epochs=50,
    imgsz=640,
    batch=16,
    device=0  # GPU ID
)

Train via CLI

yolo detect train data=/path/to/data.yaml model=yolov8n.pt epochs=50 imgsz=640 batch=16 device=0

After training, your weights will be at:

runs/detect/trainX/weights/best.pt

Update the object_detector.py model path to this file.

---

⚡ Detection Optimization

Confidence Threshold (conf)

Increase to reduce false positives.

results = self.model.predict(source=color_img, conf=0.6, verbose=False)

Image Size

Larger sizes improve accuracy, but increase inference time.

Model Variant

• yolov8n.pt → fastest, least accurate.
• yolov8s.pt → balanced.
• yolov8m.pt → most accurate, slower.

GPU Acceleration

export CUDA_VISIBLE_DEVICES=0

Class Filtering

if cls_name not in ["target_class"]:
    return

---

🧪 Testing Your Model

After swapping the model:

ros2 run object_detector yolo_detector

Look for:

✅ YOLO detector node initialized with synchronized inputs.
[📍 DETECTED] target_class 2D(320,240) 3D(0.42, 0.15, 0.80)

Verify:

• Bounding boxes appear correctly in /yolo/annotated.
• /target_pose publishes when your target is detected.

Pro Tip

If detection is unstable, collect more diverse training images with different lighting, angles, and distances.

---

Next: 📍 Customizing Waypoints