Automated vision-based dimensional inspection in car body assembly line with image sequence processing

Precise gap and flush measurements are crucial for car body assembly quality and performance. This paper proposes an automated vision-based dimensional inspection framework (VDIF) that accurately localises measurement points and computes gap/flush dimensions after creating templates defining the measurement points. By analysing image sequences from a single low-cost camera, the approach mimics a multi-view system and enables real-time, continuous inspection – an efficient alternative to manual measurement while enhancing quality control. The framework includes image filtering, measurement area detection, precise measurement point localisation, and gap/flush prediction. A pretrained ResNet model, combined with motion detection, filters irrelevant images. A multi-template matching approach based on Histogram of Oriented Gradients (HOG) is introduced to detect measurement areas, followed by a novel method for localising measurement points across different viewpoints and lighting conditions. A second-order polynomial regression model maps pixel distances from various views to real-world dimensions, dynamically estimating gap and flush using spatial relationships with adjacent points for enhanced precision. The proposed approach is validated on both synthetic and real-world assembly line data, demonstrating strong performance in image filtering, measurement detection, and gap and flush estimation. Benchmarking against alternative methods confirms its superior accuracy and efficiency, making it a scalable solution for automated real-time inspection.