PERFORMANCE EVALUATION AND PRACTICAL IMPLEMENTATION OF A DEEP LEARNING-BASED VISION SYSTEM FOR VEHICLE WHEEL ALIGNMENT

Abstract

Wheel alignment is a critical aspect of vehicle maintenance that significantly influences road safety, steering stability, fuel efficiency, tire longevity, and overall vehicle performance. Conventional wheel alignment technologies, including mechanical, laser-based, Charge-Coupled Device (CCD), and three-dimensional (3D) camera systems, provide reliable measurements but are often expensive, require specialized hardware, frequent calibration, and skilled technicians. These limitations reduce their accessibility, particularly for small and medium-sized automotive workshops and educational institutions. This study proposes the design and implementation of a deep learning-based vision system for automated vehicle wheel alignment using computer vision and artificial intelligence techniques. The proposed framework integrates digital image acquisition, camera calibration, image preprocessing, object detection, feature extraction, and geometric measurement algorithms to estimate principal wheel alignment parameters, including camber, caster, and toe angles, without the need for wheel-mounted sensors or reflective targets. An experimental research design was adopted, comprising system development and performance evaluation phases. Vehicle wheel images were collected from multiple passenger vehicles under varying lighting conditions, camera viewpoints, and workshop environments. The dataset underwent preprocessing through image resizing, normalization, augmentation, and annotation before training a deep learning object detection model. Camera calibration techniques were employed to correct lens distortion and establish accurate geometric relationships between image coordinates and real-world measurements. The trained model detected wheel features, extracted rim centers and reference points, and estimated alignment angles, which were subsequently compared with measurements obtained from a commercial wheel alignment machine. The system also achieved real-time processing capability, requiring approximately 100 ms per image and operating at about 30 frames per second on a GPU-enabled platform, while maintaining moderate computational resource utilization. Performance evaluation under varying lighting conditions, camera distances, wheel sizes, and vehicle categories confirmed the robustness and generalization capability of the proposed approach. The study demonstrates that deep learning and computer vision can provide an accurate, low-cost, and automated alternative to conventional wheel alignment systems. By reducing dependence on proprietary hardware and minimizing human intervention, the proposed system offers improved accessibility, diagnostic consistency, and operational efficiency.