Custom Object Detection & Tracking: Fine-Tuning YOLOv8 for Edge Robotics

Fine-Tuning YOLOv8 for Real-Time Conveyor Belt Tracking

The end goal of this project is to feed a robotic arm: a vision system that can detect, classify, and track custom wooden objects moving on a physical conveyor belt, at real-time inference speeds, and with sufficient positional accuracy to guide pick-and-place operations. The conveyor belt system is fully 3D printed in PETG, utilizing a modular, open-source design from MakerWorld.

The target objects for recognition are the wooden game pieces from the Catan board game (see reference image below). These four custom object classes — Church, House, Bar, and Dice — are non-standard, small-scale, and visually similar in certain lighting conditions. No existing pre-trained model has ever seen them. This rules out any zero-shot approach and demands a purpose-built, fine-tuned model.

Zero-Shot Limitations of Pre-Trained YOLOv8

The first step of any robust ML project is establishing a baseline — understanding exactly what the off-the-shelf model cannot do. Running the base YOLOv8n (Nano) model on our conveyor belt scene produced spectacularly wrong results.

The model — trained on COCO's 80 generic classes — had no concept of a "Church" shaped wooden block. Instead, it mapped visual features to the nearest object in its training distribution, confidently misclassifying the wooden shapes as "scissors," "surfboards," and "knives." A human hand in the scene was tagged as a "person." Bounding boxes were erratic, appearing and disappearing frame-to-frame with no temporal consistency.

Building a Custom Dataset from Scratch

With no existing dataset for these objects, the entire training corpus had to be built from scratch. The approach was pragmatic: capture video of the actual physical setup — same lighting, same camera angle, same conveyor belt surface — and extract frames at a rate that captures sufficient class diversity without redundancy.

Frames were extracted from a recording of the actual conveyor belt scene. This ensures the model trains on the exact domain it will be deployed in — a critical practice that eliminates domain shift between training and inference.

The raw frames were imported into Roboflow, where each object instance was manually annotated by drawing precise bounding boxes and assigning class labels. Each annotation becomes a ground truth data point that the model learns to replicate.

Simulating Real-World Physical Constraints

A fundamental challenge in custom dataset training is overfitting: a small dataset causes the model to memorize training examples rather than learning generalizable features. The solution is data augmentation — synthetically expanding the dataset by applying controlled transformations.

To ensure robust generalizations, the augmentation and preprocessing strategy was configured directly in Roboflow to mirror the specific physical constraints of the physical conveyor belt environment:

• Dataset Split — 94% Train (129 images) and 6% Validation (8 images), structured to optimize validation signals for custom tracking.
• Preprocessing — Auto-Orient applied, and all images stretched to a uniform 640x640 canvas matching YOLO's native size.
• Outputs per training example — 3x expansion factor, generating synthetic variations to multiply the training distribution.
• Rotation (Between -15° and +15°) — Simulates arbitrary orientations of wooden blocks as they are placed or roll slightly onto the moving belt.
• Brightness (Between -25% and +25%) — Models physical changes in ambient light in different environments and times of day.
• Blur (Up to 1.5px) — Accounts for motion blur introduced by the physical velocity of the conveyor belt.
• Noise (Up to 1.49% of pixels) — Replicates sensor gain and camera artifacts to make the model invariant to hardware noise.

Fine-Tuning with Ultralytics YOLOv8

Training was executed on Google Colab using a GPU runtime, leveraging the Ultralytics YOLOv8 framework. The chosen architecture was YOLOv8 Nano (YOLOv8n) — the smallest model in the YOLO8 family, deliberately selected to maximize inference speed on the target edge hardware while still achieving sufficient accuracy for the task.

# Fine-tuning command (Ultralytics CLI)
yolo task=detect mode=train \
    model=yolov8n.pt \
    data=dataset.yaml \
    epochs=100 \
    imgsz=640 \
    batch=16 \
    patience=20

The training curves tell a clear story: both training and validation losses (box, classification, DFL) converge rapidly within the first 20 epochs and stabilize with no signs of divergence or overfitting. Precision and Recall climb to near-ceiling values. The mAP@50 reaches ~97% and mAP@50-95 exceeds 90%, indicating robust detection across a range of IoU thresholds.

Real-Time Inference with Persistent Object IDs

The fine-tuned .pt weight file was deployed locally using OpenCV for camera feed capture and the Ultralytics inference API for frame-by-frame detection. The key upgrade over basic detection is the addition of a multi-object tracking algorithm.

Rather than detecting objects independently in each frame (which produces flickering boxes and no temporal consistency), a tracking algorithm like ByteTrack assigns a persistent tracking ID to each detected object across frames. This means the system knows that object #3 (a "House") at position X in frame 150 is the same object at position X+5 in frame 151 — enabling trajectory prediction and confident hand-off to the robotic arm controller.

# Inference with tracking
from ultralytics import YOLO
import cv2

model = YOLO("best.pt")  # fine-tuned weights
cap = cv2.VideoCapture(0)  # webcam / camera feed

while cap.isOpened():
    ret, frame = cap.read()
    results = model.track(frame, persist=True, tracker="bytetrack.yaml")
    annotated = results[0].plot()
    cv2.imshow("Conveyor Tracking", annotated)