Skip to content

TRAINING_GUIDE.md

Latest commit

 

History

History
74 lines (58 loc) · 2.91 KB

TRAINING_GUIDE.md

File metadata and controls

74 lines (58 loc) · 2.91 KB

🎓 Training YOLO11n for Maximum Efficiency

To get the best performance (speed + accuracy) for your specific hallway, you should fine-tune the model. Using the default model finds everything (cars, dogs, umbrellas), but you only care about people.

🚀 Strategy: "Transfer Learning"

We will teach the model to ignore everything except humans and focus specifically on the camera angle of your hallway.

Benefits

  1. Higher Accuracy: The model learns your specific lighting and background.
  2. Faster Post-Processing: By training for only 1 class ('person'), the Pi does less math filtering out other objects.

🛠️ Step-by-Step Guide

Phase 1: Data Collection (On Raspberry Pi)

Capture images from your actual deployed camera.

  1. Create a folder: mkdir -p dataset/images
  2. Run a simple script to save a frame every 5 seconds (or use crowd_counting_rpi.py and manually save interesting frames).
  3. Aim for: 100-200 images of people walking in your hallway. Varied clothing and lighting helps.

Phase 2: Annotation (On PC)

Label your images.

  1. Move images to your PC.
  2. Use a tool like Roboflow (easiest, web-based) or CVAT.
  3. Draw boxes around every person.
  4. Export Dataset: Choose "YOLOv8" format (compatible with YOLO11).

Phase 3: Training (On Google Colab / PC with GPU)

⚠️ DO NOT TRAIN ON RASPBERRY PI. It is too slow. Use a Cloud GPU (free).

  1. Open Google Colab.
  2. Select Runtime > Change runtime type > T4 GPU.
  3. Run this code:
# 1. Install Ultralytics
!pip install ultralytics

# 2. Upload your dataset zip file from Roboflow/CVAT
# (Unzip it into a folder called 'datasets')

# 3. Train the model
from ultralytics import YOLO

# Load the nano model
model = YOLO('yolo11n.pt') 

# Train
results = model.train(
    data='/content/datasets/data.yaml',  # Path to your dataset config
    epochs=50,                           # 50-100 is usually enough for fine-tuning
    imgsz=640,                           # Maintain resolution
    device=0,                            # Use GPU
    project='hallway_monitor',
    name='v1'
)

Phase 4: Deploy Back to Pi

  1. Download the trained best.pt file from Colab.
  2. Transfer it to your Raspberry Pi.
  3. Export for NCS2 (Critical Step): 
    yolo export model=best.pt format=openvino half=True
  4. Update your crowd_counting_rpi.py to use your new model path or pass it via command line: 
    python3 crowd_counting_rpi.py --model best_openvino_model/

📉 Advanced: Reducing Model Size (Quantization)

Since you are using Intel NCS2, you are already using FP16 (Half Precision). For further speed (at cost of some accuracy), you can explore INT8 Quantization with the OpenVINO toolkit, but this is complex and usually not needed if you stick to the Nano (n) model.