Ensuring precise quality control in automotive manufacturing is crucial for both safety and efficiency. One of the most common yet critical defects in vehicle assembly is incorrect rim mounting, which can significantly impact the stability and handling of a car. Traditionally, this inspection process relies on human workers, who are susceptible to fatigue and errors, leading to inconsistent results. Our project introduces a real-time, fully automated system that detects, classifies, and verifies the rims on all four wheels of a vehicle. By leveraging a combination of advanced computer vision techniques and deep learning models, we enhance the speed, accuracy, and reliability of the inspection process, reducing the need for manual intervention while ensuring high precision in automotive quality control.
Advanced AI-Powered Detection and Classification
The system integrates state-of-the-art object detection and classification techniques to automatically identify and analyze wheels in real time. Using a two-stage convolutional neural network (CNN) pipeline, the first model detects cars and their wheels, while the second model performs rim classification and bolt detection. Unlike traditional methods such as the Hough Transform, which struggle with occlusions and variations in lighting, our AI-driven approach adapts to diverse manufacturing environments and maintains a consistently high detection accuracy. The classification module, trained on a dataset of thousands of rim images, can differentiate between multiple rim types and detect discrepancies in real-time, ensuring that all wheels on a vehicle match in size and design.
High-Precision Rim Size Estimation and Tracking
One of the key innovations of our system is its ability to estimate the real dimensions of the rim with high precision. By utilizing a combination of bolt detection and ellipse fitting techniques, the system accurately determines the rim’s size relative to a known reference—the pitch circle diameter of the wheel bolts. This method eliminates the need for complex calibration and provides an automated way to validate rim dimensions within an industrial production line. Additionally, the system employs real-time object tracking to maintain consistent identification of each wheel throughout the inspection process. This ensures that classification results remain stable across multiple frames, minimizing errors caused by transient visual obstructions or lighting fluctuations.
|
WHEEL22 dataset: 1+21 rim categories. C00 category is used for handling occlusions. |
Publicly Available Datasets for Research and Development
To facilitate further advancements in automotive computer vision, we have compiled and publicly released three comprehensive datasets: CWD1500 for car and wheel detection, WHEEL22 for rim classification, and RB600 for bolt detection. These datasets, collected in a real-world manufacturing environment, provide high-quality labeled images that can be used for training and benchmarking machine learning models. By making these datasets freely accessible for research purposes, we aim to support the broader AI community in developing more efficient and scalable solutions for automated quality control in the automotive industry.
Dataset: HERE on Kagggle
Contact person: Adam Novozámský
