frmc/Truck-Wheel-Rim-Line-Dataset

Vision-Based Non-Contact Weigh-in-Motion (WIM) Perception Dataset

📖 Background & Motivation

Traditional contact-based Weigh-in-Motion (WIM) systems (like quartz piezoelectric sensors) have practical limitations: they are expensive, prone to wear and tear under heavy loads, and often cause traffic bottlenecks at toll stations.

To address this, we created this dataset to help build a non-contact dynamic weighing system using computer vision. By tracking highly visible rigid features on moving trucks (such as wheel rims, axles, and chassis baselines), we can estimate the suspension deformation and calculate the vehicle's weight. This approach serves as a low-cost, efficient pre-screening tool for highway overload enforcement.

Note: This repository contains a curated subset of 353 high-quality images from our larger WIM dataset, collected at actual highway toll stations under various real-world conditions.

🚀 Key Features

This dataset is designed to train perception models that tackle two main challenges in visual WIM measurement:

1. Scale Calibration: Instead of relying on fragile external calibration boards, we use the vehicle's wheel rim. By fitting an ellipse to the rim mask and extracting its minor axis (which is largely unaffected by perspective distortion), we can accurately convert 2D pixel measurements into 3D physical space in real time.
2. Handling Occlusion: Directly observing truck suspensions is difficult due to mud and structural interference. To work around this, our annotations focus on visible rigid parts (the axle center, rim, and chassis line) to indirectly calculate the hidden suspension compression.

📂 Dataset Structure

The dataset contains a single pool of 353 images. We deliberately left it undivided so you can set up custom k-fold cross-validation or specific train/val splits based on your own research requirements.

• Pose/ : Contains images and YOLOv8-Pose labels for axle keypoint detection.
• Segmentation/ : Contains the same images and YOLOv8-Seg labels for rim segmentation and chassis line detection.

Task 1: Axle Keypoint Detection (Pose)

The goal here is to locate the center of the wheel axle, which helps counter the motion blur caused by high-speed vehicles.

• Class: 0: wheel
• Format: Standard YOLO Pose format (class x_center y_center width height px py visibility)

Task 2: Rim & Chassis Baseline Extraction (Segmentation & Detection)

The goal is to extract a precise mask of the metal rim for geometric calibration, and to detect the flat bottom edge of the cargo box to serve as a horizontal reference line.

• Classes: - 0: rim (Instance Segmentation using polygon coordinates)
  • 1: line (Object Detection for the chassis baseline using standard bounding boxes)
• Format: Standard YOLO Segmentation format.

💻 How to Use (Ultralytics YOLO)

If you want to train on the entire dataset at once, just point both train and val to the same images directory in your .yaml configuration files.

Example: Training the Pose Model

yolo task=pose mode=train data=Pose/pose.yaml model=yolov8n-pose.pt epochs=100 imgsz=640